Adding codebeaver.yml

.github/workflows/release.yml

@@ -9,7 +9,7 @@ on:
       - 0.*
 
 env:
-  POETRY_VERSION: "1.8.5"
+  POETRY_VERSION: "1.7.1"
 
 jobs:
   if_release:

.github/workflows/test.yml

@@ -7,7 +7,7 @@ on:
     branches: [main]
 
 env:
-  POETRY_VERSION: "1.8.5"
+  POETRY_VERSION: "1.7.1"
   OPENAI_API_KEY: "sk-fake"
 
 jobs:

.gitignore

@@ -17,5 +17,3 @@ inv/
 scripts/
 docx/
 agentic_security.toml
-/venv
-*.csv

Dockerfile

@@ -1,5 +1,5 @@
 # Build stage
-FROM python:3.11-slim AS builder
+FROM python:3.11-slim as builder
 
 WORKDIR /app
 
@@ -14,15 +14,8 @@ RUN poetry self add "poetry-plugin-export"
 # Copy only dependency files to leverage Docker layer caching
 COPY pyproject.toml poetry.lock ./
 
-# update lock file to avoid failure
-RUN poetry lock
-
 # Install dependencies
 RUN poetry export -f requirements.txt --without-hashes -o requirements.txt
-
-# Install wheel (required to build packages like fire)
-RUN pip install --upgrade pip setuptools wheel
-
 RUN pip install --no-cache-dir -r requirements.txt
 
 # Runtime stage

Readme.md

@@ -21,7 +21,9 @@
   <a href="https://pypi.org/project/agentic-security/">
     <img alt="PyPI Version" src="https://img.shields.io/pypi/v/agentic-security?style=for-the-badge&logo=pypi&labelColor=000000&color=00CCFF" />
   </a>
-
+  <a href="https://discord.gg/stw3DfZQ">
+    <img alt="Join Discord" src="https://img.shields.io/badge/Discord-Join%20Us-black?style=for-the-badge&logo=discord&labelColor=000000&color=DD55FF" />
+  </a>
 </p>
 
 
@@ -81,7 +83,7 @@ agentic_security --port=PORT --host=HOST
 
 ## UI 🧙
 
-<img width="100%" alt="booking-screen" src="https://raw.githubusercontent.com/msoedov/agentic_security/refs/heads/main/docs/images/demo.gif">
+<img width="100%" alt="booking-screen" src="https://res.cloudinary.com/dq0w2rtm9/image/upload/v1736433557/z0bsyzhsqlgcr3w4ovwp.gif">
 
 ## LLM kwargs
 
@@ -400,31 +402,16 @@ This setup ensures a continuous integration approach towards maintaining securit
 
 The `Module` class is designed to manage prompt processing and interaction with external AI models and tools. It supports fetching, processing, and posting prompts asynchronously for model vulnerabilities. Check out [module.md](https://github.com/msoedov/agentic_security/blob/main/docs/module.md) for details.
 
-
-## MCP server
-
-```shell
-pip install -U mcp
-
-# From cloned directory
-mcp install agentic_security/mcp/main.py
-```
-
 ## Documentation
 
 For more detailed information on how to use Agentic Security, including advanced features and customization options, please refer to the official documentation.
 
 ## Roadmap and Future Goals
 
-
-
-We’re just getting started! Here’s what’s on the horizon:
-
-- **RL-Powered Attacks**: An attacker LLM trained with reinforcement learning to dynamically evolve jailbreaks and outsmart defenses.
-- **Massive Dataset Expansion**: Scaling to 100,000+ prompts across text, image, and audio modalities—curated for real-world threats.
-- **Daily Attack Updates**: Fresh attack vectors delivered daily, keeping your scans ahead of the curve.
-- **Community Modules**: A plug-and-play ecosystem where you can share and deploy custom probes, datasets, and integrations.
-
+- \[ \] Expand dataset variety
+- \[ \] Introduce two new attack vectors
+- \[ \] Develop initial attacker LLM
+- \[ \] Complete integration of OWASP Top 10 classification
 
 | Tool                    | Source                                                                        | Integrated |
 |-------------------------|-------------------------------------------------------------------------------|------------|
@@ -436,7 +423,6 @@ We’re just getting started! Here’s what’s on the horizon:
 
 Note: All dates are tentative and subject to change based on project progress and priorities.
 
-
 ## 👋 Contributing
 
 Contributions to Agentic Security are welcome! If you'd like to contribute, please follow these steps:
@@ -453,9 +439,4 @@ Before contributing, please read the contributing guidelines.
 
 Agentic Security is released under the Apache License v2.
 
-
-## 🚫 No Cryptocurrency Affiliation
-
-Agentic Security is focused solely on AI security and has no affiliation with cryptocurrency projects, blockchain technologies, or related initiatives. Our mission is to advance the safety and reliability of AI systems—no tokens, no coins, just code.
-
 ## Contact us

agentic_security/__init__.py

@@ -1,3 +1,3 @@
-from .lib import SecurityScanner
+from .lib import AgenticSecurity
 
-__all__ = ["SecurityScanner"]
+__all__ = ["AgenticSecurity"]

agentic_security/__main__.py

@@ -5,7 +5,7 @@ import fire
 import uvicorn
 
 from agentic_security.app import app
-from agentic_security.lib import SecurityScanner
+from agentic_security.lib import AgenticSecurity
 from agentic_security.misc.banner import init_banner
 
 
@@ -32,14 +32,14 @@ class CLI:
         Run Agentic Security in CI mode.
         """
         sys.path.append(os.path.dirname("."))
-        SecurityScanner().entrypoint()
+        AgenticSecurity().entrypoint()
 
     def init(self, host: str = "0.0.0.0", port: int = 8718):
         """
         Generate the default CI configuration file.
         """
         sys.path.append(os.path.dirname("."))
-        SecurityScanner().generate_default_settings(host, port)
+        AgenticSecurity().generate_default_settings(host, port)
 
     i = init
 
@@ -48,7 +48,7 @@ class CLI:
         List all available security checks.
         """
         sys.path.append(os.path.dirname("."))
-        SecurityScanner().list_checks()
+        AgenticSecurity().list_checks()
 
 
 def main():

agentic_security/agents/operator_crew.py

@@ -1,4 +1,5 @@
 import asyncio
+import logging
 import os
 from typing import Any
 
@@ -9,11 +10,12 @@ from pydantic import BaseModel, ConfigDict, Field
 
 # Assuming LLMSpec is defined elsewhere; placeholder import
 from agentic_security.http_spec import LLMSpec
-from agentic_security.logutils import logger
 
 LLM_SPECS = []  # Populate with LLM spec strings if needed
 
 # Configure logging
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
 
 
 # Define AgentSpecification model

agentic_security/agents/operator_pydantic.py

@@ -1,4 +1,5 @@
 import asyncio
+import logging
 from typing import Any
 
 import httpx
@@ -7,10 +8,13 @@ from pydantic_ai import Agent, RunContext, Tool
 
 # Assuming LLMSpec is defined elsewhere; placeholder import
 from agentic_security.http_spec import LLMSpec
-from agentic_security.logutils import logger
 
 LLM_SPECS = []  # Populate this list with LLM spec strings if needed
 
+# Configure logging
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
 
 # Define AgentSpecification model
 class AgentSpecification(BaseModel):

agentic_security/config.py

@@ -1,10 +1,9 @@
 from functools import lru_cache
 
 import tomli
+from loguru import logger
 
-from agentic_security.logutils import logger
-
-SETTINGS_VERSION = 2
+SETTINGS_VERSION = 1
 
 
 @lru_cache(maxsize=1)
@@ -143,13 +142,6 @@ use_disk_cache = false
 retry = 3
 timeout_connect = 30
 timeout_response = 90
-
-[fuzzer]
-max_prompt_lenght = 2048
-budget_multiplier = 100000000
-initial_optimizer_points = 25
-min_failure_samples = 5
-failure_rate_threshold = 0.5
 """.replace(
                     "$HOST", host
                 )

agentic_security/core/app.py

@@ -4,14 +4,10 @@ from asyncio import Event, Queue
 from fastapi import FastAPI
 from fastapi.responses import ORJSONResponse
 
-from agentic_security.http_spec import LLMSpec
-
 tools_inbox: Queue = Queue()
 stop_event: Event = Event()
 current_run: str = {"spec": "", "id": ""}
-_secrets: dict[str, str] = {}
-
-current_run: dict[str, int | LLMSpec] = {"spec": "", "id": ""}
+_secrets = {}
 
 
 def create_app() -> FastAPI:
@@ -30,29 +26,29 @@ def get_stop_event() -> Event:
     return stop_event
 
 
-def get_current_run() -> dict[str, int | LLMSpec]:
+def get_current_run() -> str:
     """Get the current run id."""
     return current_run
 
 
-def set_current_run(spec: LLMSpec) -> dict[str, int | LLMSpec]:
+def set_current_run(spec):
     """Set the current run id."""
     current_run["id"] = hash(id(spec))
     current_run["spec"] = spec
     return current_run
 
 
-def get_secrets() -> dict[str, str]:
+def get_secrets():
     return _secrets
 
 
-def set_secrets(secrets: dict[str, str]) -> dict[str, str]:
+def set_secrets(secrets):
     _secrets.update(secrets)
     expand_secrets(_secrets)
     return _secrets
 
 
-def expand_secrets(secrets: dict[str, str]) -> None:
+def expand_secrets(secrets):
     for key in secrets:
         val = secrets[key]
         if val.startswith("$"):

agentic_security/core/logging.py

@@ -1,5 +1,26 @@
-from agentic_security.logutils import set_log_level_to_info
+from logging import config
 
 
 def setup_logging():
-    return set_log_level_to_info()
+    config.dictConfig(
+        {
+            "version": 1,
+            "disable_existing_loggers": True,
+            "handlers": {
+                "console": {
+                    "class": "logging.StreamHandler",
+                },
+            },
+            "root": {
+                "handlers": ["console"],
+                "level": "INFO",
+            },
+            "loggers": {
+                "uvicorn.access": {
+                    "level": "ERROR",  # Set higher log level to suppress info logs globally
+                    "handlers": ["console"],
+                    "propagate": False,
+                }
+            },
+        }
+    )

agentic_security/http_spec.py

@@ -1,6 +1,5 @@
 import base64
 from enum import Enum
-from urllib.parse import urlparse
 
 import httpx
 from pydantic import BaseModel
@@ -160,14 +159,6 @@ def parse_http_spec(http_spec: str) -> LLMSpec:
     # Extract the method and URL from the first line
     method, url = lines[0].split(" ")[0:2]
 
-    # Check url validity
-    valid_url = urlparse(url)
-    # if missing the correct formatting ://, urlparse.netloc will be empty
-    if valid_url.scheme not in ("http", "https") or not valid_url.netloc:
-        raise InvalidHTTPSpecError(
-            f"Invalid URL: {url}. Ensure it starts with 'http://' or 'https://'"
-        )
-
     # Initialize headers and body
     headers = {}
     body = ""

agentic_security/lib.py

@@ -4,12 +4,12 @@ from datetime import datetime
 
 import colorama
 import tqdm.asyncio
+from loguru import logger
 from rich.console import Console
 from rich.table import Table
 from tabulate import tabulate
 
 from agentic_security.config import SettingsMixin  # Importing the configuration mixin
-from agentic_security.logutils import logger
 from agentic_security.primitives import Scan
 from agentic_security.probe_data import REGISTRY
 from agentic_security.routes.scan import streaming_response_generator
@@ -23,7 +23,7 @@ YELLOW = colorama.Fore.YELLOW
 BLUE = colorama.Fore.BLUE
 
 
-class SecurityScanner(SettingsMixin):
+class AgenticSecurity(SettingsMixin):
     @classmethod
     async def async_scan(
         cls,
@@ -202,7 +202,7 @@ class SecurityScanner(SettingsMixin):
         self.load_config(self.default_path)
         logger.info("Configuration loaded successfully.")
         print(self.config)
-        datasets = list(self.get_config_value("modules", {}).values())
+        datasets = list(self.get_config_value("modules").values())
         for d in datasets:
             d["selected"] = True
         self.scan(

agentic_security/logutils.py

@@ -1,146 +0,0 @@
-# import sys
-
-# from loguru import logger
-
-# # Define custom colors
-# BLUE = "#89CFF0"
-# BROWN = "#8B4513"  # Brown for DEBUG
-
-# # Define custom log level colors
-# logger.level("DEBUG", color=f"<fg {BROWN}>")
-# logger.level("INFO", color=f"<fg {BLUE}>")
-
-# # Define custom log format with aligned messages and colored levels
-# LOG_FORMAT = (
-#     "<level>{level:<8}</level> "  # Properly formatted and colored log level
-#     "<level>{message:<100}</level> "  # Left-aligned message for readability
-#     "<cyan>{file.name}</cyan>:<cyan>{line}</cyan>"  # File name and line number in cyan
-# )
-
-# # Remove default handlers and add a new one with custom formatting
-# logger.remove()
-# logger.add(sys.stdout, format=LOG_FORMAT, level="DEBUG", colorize=True)
-import logging
-import logging.config
-import time
-from collections.abc import Callable, Coroutine
-from functools import wraps
-from os import getenv
-from typing import Any, ParamSpec, TypeVar
-
-LOGGER_NAME = None
-
-LOGGING_CONFIG = {
-    "version": 1,
-    "disable_existing_loggers": False,
-    "formatters": {
-        "rich": {"format": "%(message)s", "datefmt": "[%X]"},
-    },
-    "handlers": {
-        "rich": {
-            "class": "rich.logging.RichHandler",
-            "level": "INFO",
-            "formatter": "rich",
-            "show_time": False,
-            "rich_tracebacks": False,
-            "show_path": lambda: True if getenv("API_RUNTIME") == "dev" else False,
-            "tracebacks_show_locals": False,
-        },
-    },
-    "loggers": {
-        "": {  # Root logger configuration
-            "level": "INFO",
-            "handlers": ["rich"],
-            "propagate": True,
-        },
-        "httpx": {  # Disable httpx logging
-            "level": "WARNING",  # Suppress DEBUG and INFO messages from httpx
-            "handlers": [],
-            "propagate": False,
-        },
-        "uvicorn.access": {  # Disable uvicorn.access logging
-            "level": "WARNING",  # Suppress DEBUG and INFO messages from uvicorn.access
-            "handlers": [],
-            "propagate": False,
-        },
-    },
-}
-
-
-def configure_logging():
-    # Apply the dictionary configuration
-    logging.config.dictConfig(LOGGING_CONFIG)
-
-    # Get and return the logger
-    logger = logging.getLogger(LOGGER_NAME)
-    return logger
-
-
-logger: logging.Logger = configure_logging()
-
-
-def set_log_level_to_debug():
-    logger = logging.getLogger(LOGGER_NAME)
-    logger.setLevel(logging.DEBUG)
-    # Update handler level as well
-    for handler in logger.handlers:
-        handler.setLevel(logging.DEBUG)
-
-
-def set_log_level_to_info():
-    logger = logging.getLogger(LOGGER_NAME)
-    logger.setLevel(logging.INFO)
-    # Update handler level as well
-    for handler in logger.handlers:
-        handler.setLevel(logging.INFO)
-
-
-# Set initial log level
-set_log_level_to_info()
-
-
-# Define generic type variables for return type and parameters
-R = TypeVar("R")
-P = ParamSpec("P")
-
-
-def time_execution_sync(
-    additional_text: str = "",
-) -> Callable[[Callable[P, R]], Callable[P, R]]:
-    def decorator(func: Callable[P, R]) -> Callable[P, R]:
-        @wraps(func)
-        def wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
-            start_time = time.time()
-            result = func(*args, **kwargs)
-            execution_time = time.time() - start_time
-            logger.debug(
-                f"{additional_text} Execution time: {execution_time:.2f} seconds"
-            )
-            return result
-
-        return wrapper
-
-    return decorator
-
-
-def time_execution_async(
-    additional_text: str = "",
-) -> Callable[
-    [Callable[P, Coroutine[Any, Any, R]]], Callable[P, Coroutine[Any, Any, R]]
-]:
-    def decorator(
-        func: Callable[P, Coroutine[Any, Any, R]]
-    ) -> Callable[P, Coroutine[Any, Any, R]]:
-        @wraps(func)
-        async def wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
-            start_time = time.time()
-            result = await func(*args, **kwargs)
-            execution_time = time.time() - start_time
-            logger.debug(
-                f"{additional_text} Execution time: {execution_time:.2f} seconds"
-            )
-            return result
-
-        return wrapper
-
-    return decorator

agentic_security/mcp/client.py

@@ -1,54 +0,0 @@
-import asyncio
-
-from mcp import ClientSession, StdioServerParameters
-from mcp.client.stdio import stdio_client
-
-# Create server parameters for stdio connection
-server_params = StdioServerParameters(
-    command="python",  # Executable
-    args=["agentic_security/mcp/main.py"],  # Your server script
-    env=None,  # Optional environment variables
-)
-
-
-async def run() -> None:
-    async with stdio_client(server_params) as (read, write):
-        async with ClientSession(read, write) as session:
-            # Initialize the connection --> connection does not work
-            await session.initialize()
-
-            # List available prompts, resources, and tools --> no avalialbe tools
-            prompts = await session.list_prompts()
-            print(f"Available prompts: {prompts}")
-
-            resources = await session.list_resources()
-            print(f"Available resources: {resources}")
-
-            tools = await session.list_tools()
-            print(f"Available tools: {tools}")
-
-            # Call the echo tool --> echo tool iisue
-            echo_result = await session.call_tool(
-                "echo_tool", arguments={"message": "Hello from client!"}
-            )
-            print(f"Tool result: {echo_result}")
-
-            # # Read the echo resource
-            # echo_content, mime_type = await session.read_resource(
-            #     "echo://Hello_resource"
-            # )
-            # print(f"Resource content: {echo_content}")
-            # print(f"Resource MIME type: {mime_type}")
-
-            # # Get and use the echo prompt
-            # prompt_result = await session.get_prompt(
-            #     "echo_prompt", arguments={"message": "Hello prompt!"}
-            # )
-            # print(f"Prompt result: {prompt_result}")
-
-            # You can perform additional operations here as needed
-            return prompts, resources, tools
-
-
-if __name__ == "__main__":
-    asyncio.run(run())

agentic_security/mcp/main.py

@@ -1,109 +0,0 @@
-import httpx
-from mcp.server.fastmcp import FastMCP
-
-# Initialize MCP server
-mcp = FastMCP(
-    name="Agentic Security MCP Server",
-    description="MCP server to interact with LLM scanning test",
-    dependencies=["httpx"],
-)
-
-# FastAPI Server Configuration
-AGENTIC_SECURITY = "http://0.0.0.0:8718"
-
-
-@mcp.tool()
-async def verify_llm(spec: str) -> dict:
-    """
-    Verify an LLM model specification using the FastAPI server
-
-    Returns:
-        dict: containing the verification result form the FastAPI server
-
-    Args: spect(str):  The specification of the LLM model to verify.
-
-    """
-    url = f"{AGENTIC_SECURITY}/verify"
-    async with httpx.AsyncClient() as client:
-        response = await client.post(url, json={"spec": spec})
-        return response.json()
-
-
-@mcp.tool()
-async def start_scan(
-    llmSpec: str,
-    maxBudget: int,
-    optimize: bool = False,
-    enableMultiStepAttack: bool = False,
-) -> dict:
-    """
-    Start an LLM security scan via the FastAPI server.
-    Returns:
-        dict: The scan initiation result from the FastAPI server.
-
-    Args:
-        llmSpec (str): The specification of the LLM model.
-        maxBudget (int): The maximum budget for the scan.
-        optimize (bool, optional): Whether to enable optimization during scanning. Defaults to False.
-        enableMultiStepAttack (bool, optional): Whether to enable multi-step attack
-
-    """
-    url = f"{AGENTIC_SECURITY}/scan"
-    payload = {
-        "llmSpec": llmSpec,
-        "maxBudget": maxBudget,
-        "datasets": [],
-        "optimize": optimize,
-        "enableMultiStepAttack": enableMultiStepAttack,
-        "probe_datasets": [],
-        "secrets": {},
-    }
-    async with httpx.AsyncClient() as client:
-        response = await client.post(url, json=payload)
-        return response.json()
-
-
-@mcp.tool()
-async def stop_scan() -> dict:
-    """Stop an ongoing scan via the FastAPI server.
-
-    Returns:
-        dict: The confirmation from the FastAPI server that the scan has been stopped.
-    """
-    url = f"{AGENTIC_SECURITY}/stop"
-    async with httpx.AsyncClient() as client:
-        response = await client.post(url)
-        return response.json()
-
-
-@mcp.tool()
-async def get_data_config() -> list:
-    """
-    Retrieve data configuration from the FastAPI server.
-
-    Returns:
-        list: The response from the FastAPI server, confirming the scan has been stopped.
-    """
-    url = f"{AGENTIC_SECURITY}/v1/data-config"
-    async with httpx.AsyncClient() as client:
-        response = await client.get(url)
-        return response.json()
-
-
-@mcp.tool()
-async def get_spec_templates() -> list:
-    """
-    Retrieve data configuration from the FastAPI server.
-
-    Returns:
-        list: The LLM specification templates from the FastAPI server.
-    """
-    url = f"{AGENTIC_SECURITY}/v1/llm-specs"
-    async with httpx.AsyncClient() as client:
-        response = await client.get(url)
-        return response.json()
-
-
-# Run the MCP server
-if __name__ == "__main__":
-    mcp.run()

agentic_security/middleware/logging.py

@@ -1,8 +1,7 @@
 from fastapi import Request
+from loguru import logger
 from starlette.middleware.base import BaseHTTPMiddleware
 
-from agentic_security.logutils import logger
-
 
 class LogNon200ResponsesMiddleware(BaseHTTPMiddleware):
     async def dispatch(self, request: Request, call_next):

agentic_security/misc/banner.py

@@ -8,14 +8,14 @@ except ImportError:
 
 
 def generate_banner(
-    title: str = "Agentic Security",
-    font: str = "slant",
-    version: str = "v2.1.0",
-    tagline: str = "Proactive Threat Detection & Automated Security Protocols",
-    author: str = "Developed by: [Security Team]",
-    website: str = "Website: https://github.com/msoedov/agentic_security",
-    warning: str | None = "",  # Using Optional for warning since it might be None
-) -> str:
+    title="Agentic Security",
+    font="slant",
+    version="v2.1.0",
+    tagline="Proactive Threat Detection & Automated Security Protocols",
+    author="Developed by: [Security Team]",
+    website="Website: https://github.com/msoedov/agentic_security",
+    warning="",
+):
     """Generate a visually enhanced banner with dynamic width and borders."""
     # Define the text elements
 
@@ -80,13 +80,8 @@ def generate_banner(
 
 
 def init_banner():
-    return
     ver = version("agentic_security")
-    try:
-        print(generate_banner(version=ver))
-    except Exception:
-        # UnicodeEncodeError with codec on some systems
-        pass
+    print(generate_banner(version=ver))
 
 
 if __name__ == "__main__":

agentic_security/primitives/__init__.py

@@ -1,6 +1,5 @@
-# noqa
-from agentic_security.primitives.models import CompletionRequest  # noqa
 from agentic_security.primitives.models import (  # noqa
+    CompletionRequest,
     FileProbeResponse,
     LLMInfo,
     Message,

agentic_security/probe_actor/fuzzer.py

@@ -3,44 +3,32 @@ import random
 import time
 from collections.abc import AsyncGenerator
 from json import JSONDecodeError
-from typing import Any
 
 import httpx
+import pandas as pd
+from loguru import logger
 from skopt import Optimizer
 from skopt.space import Real
 
-from agentic_security.config import settings_var
 from agentic_security.http_spec import Modality
-from agentic_security.logutils import logger
 from agentic_security.primitives import Scan, ScanResult
 from agentic_security.probe_actor.cost_module import calculate_cost
 from agentic_security.probe_actor.refusal import refusal_heuristic
-from agentic_security.probe_actor.state import FuzzerState
 from agentic_security.probe_data import audio_generator, image_generator, msj_data
 from agentic_security.probe_data.data import prepare_prompts
 
-MAX_PROMPT_LENGTH = settings_var("fuzzer.max_prompt_lenght", 2048)
-BUDGET_MULTIPLIER = settings_var("fuzzer.budget_multiplier", 100000000)
-INITIAL_OPTIMIZER_POINTS = settings_var("fuzzer.initial_optimizer_points", 25)
-MIN_FAILURE_SAMPLES = settings_var("min_failure_samples", 5)
-FAILURE_RATE_THRESHOLD = settings_var("failure_rate_threshold", 0.5)
+# TODO: full log file
+
+MAX_PROMPT_LENGTH = 2048
+BUDGET_MULTIPLIER = 100_000_000
+INITIAL_OPTIMIZER_POINTS = 25
+MIN_FAILURE_SAMPLES = 5
+FAILURE_RATE_THRESHOLD = 0.5
 
 
 async def generate_prompts(
     prompts: list[str] | AsyncGenerator,
 ) -> AsyncGenerator[str, None]:
-    """
-    Asynchronously generates and yields individual prompts.
-
-    If the input is a list of strings, the function sequentially yields each string.
-    If the input is an asynchronous generator, it forwards each generated prompt.
-
-    Args:
-        prompts (list[str] | AsyncGenerator): A list of strings or an asynchronous generator of prompts.
-
-    Yields:
-        str: An individual prompt from the list or the asynchronous generator.
-    """
     if isinstance(prompts, list):
         for prompt in prompts:
             yield prompt
@@ -49,21 +37,7 @@ async def generate_prompts(
             yield prompt
 
 
-def get_modality_adapter(llm_spec):
-    """
-    Returns the appropriate request adapter based on the modality of the LLM specification.
-
-    Depending on the modality of `llm_spec`, the function selects the corresponding request adapter.
-    If the modality is IMAGE or AUDIO, it returns an adapter for handling the respective type.
-    If the modality is TEXT or an unrecognized type, it returns `llm_spec` as is.
-
-    Args:
-        llm_spec: An object containing modality information for the LLM.
-
-    Returns:
-        RequestAdapter | llm_spec: An instance of the appropriate request adapter
-        or the original `llm_spec` if no adaptation is needed.
-    """
+def multi_modality_spec(llm_spec):
     match llm_spec.modality:
         case Modality.IMAGE:
             return image_generator.RequestAdapter(llm_spec)
@@ -76,71 +50,40 @@ def get_modality_adapter(llm_spec):
 
 
 async def process_prompt(
-    request_factory,
-    prompt: str,
-    tokens: int,
-    module_name: str,
-    fuzzer_state: FuzzerState,
+    request_factory, prompt, tokens, module_name, refusals, errors, outputs
 ) -> tuple[int, bool]:
     """
-    Processes a single prompt using the provided request factory and updates tracking lists.
-
-    This function sends the given `prompt` to the `request_factory`, checks for errors, and updates
-    the `tokens`, `refusals`, `errors`, and `outputs` lists accordingly. If the request fails or
-    the response indicates a refusal, the function records the issue and returns the updated token count
-    along with a boolean indicating whether the prompt was refused.
-
-    Args:
-        request_factory: An object with a `fn` method used to send the prompt.
-        prompt (str): The input prompt to be processed.
-        tokens (int): The current token count, which will be updated.
-        module_name (str): The name of the module handling the request.
-        fuzzer_state: State tracking object for the fuzzer
-
-    Returns:
-        tuple[int, bool]: Updated token count and a boolean indicating if the prompt was refused.
+    Process a single prompt and update the token count and failure status.
     """
     try:
         response = await request_factory.fn(prompt=prompt)
-
-        # Handle HTTP errors
         if response.status_code == 422:
             logger.error(f"Invalid prompt: {prompt}, error=422")
-            fuzzer_state.add_error(module_name, prompt, 422, "Invalid prompt")
+            errors.append((module_name, prompt, 422, "Invalid prompt"))
             return tokens, True
 
         if response.status_code >= 400:
             logger.error(f"HTTP {response.status_code} {response.content=}")
-            fuzzer_state.add_error(
-                module_name, prompt, response.status_code, response.text
-            )
+            errors.append((module_name, prompt, response.status_code, response.text))
             return tokens, True
-
-        # Process successful response
         response_text = response.text
         tokens += len(response_text.split())
 
-        # Check if the response indicates a refusal
         refused = refusal_heuristic(response.json())
         if refused:
-            fuzzer_state.add_refusal(
-                module_name, prompt, response.status_code, response_text
-            )
+            refusals.append((module_name, prompt, response.status_code, response_text))
 
-        fuzzer_state.add_output(module_name, prompt, response_text, refused)
+        outputs.append((module_name, prompt, response_text, refused))
         return tokens, refused
 
     except httpx.RequestError as exc:
         logger.error(f"Request error: {exc}")
-        fuzzer_state.add_error(module_name, prompt, "?", str(exc))
+        errors.append((module_name, prompt, "?", str(exc)))
         return tokens, True
     except JSONDecodeError as json_decode_error:
-        logger.error(f"JSON error: {json_decode_error}")
-        fuzzer_state.add_error(module_name, prompt, "?", str(json_decode_error))
+        logger.error(f"Jason error: {json_decode_error}")
+        errors.append((module_name, prompt, "?", str(json_decode_error)))
         return tokens, True
-    except Exception as e:
-        logger.exception(f"Unexpected error: {e}")
-        return tokens, False
 
 
 async def process_prompt_batch(
@@ -148,29 +91,14 @@ async def process_prompt_batch(
     prompts: list[str],
     tokens: int,
     module_name: str,
-    fuzzer_state: FuzzerState,
+    refusals,
+    errors,
+    outputs,
 ) -> tuple[int, int]:
-    """
-    Processes a batch of prompts asynchronously and aggregates the results.
-
-    This function sends multiple prompts concurrently using `process_prompt`,
-    collects the token count and failure status for each prompt, and returns
-    the total number of tokens processed and the number of failed prompts.
-
-    Args:
-        request_factory: An object with a `fn` method used to send the prompts.
-        prompts (list[str]): A list of input prompts to be processed.
-        tokens (int): The initial token count, which will be updated.
-        module_name (str): The name of the module handling the request.
-        fuzzer_state: State tracking object for the fuzzer
-
-    Returns:
-        tuple[int, int]:
-            - Total number of tokens processed.
-            - Number of failed prompts.
-    """
     tasks = [
-        process_prompt(request_factory, p, tokens, module_name, fuzzer_state)
+        process_prompt(
+            request_factory, p, tokens, module_name, refusals, errors, outputs
+        )
         for p in prompts
     ]
     results = await asyncio.gather(*tasks)
@@ -179,238 +107,138 @@ async def process_prompt_batch(
     return total_tokens, failures
 
 
-async def scan_module(
-    request_factory,
-    module,
-    fuzzer_state: FuzzerState,
-    processed_prompts: int = 0,
-    total_prompts: int = 0,
-    max_budget: int = 0,
-    total_tokens: int = 0,
-    optimize: bool = False,
-    stop_event: asyncio.Event | None = None,
-) -> AsyncGenerator[dict[str, Any], None]:
-    """
-    Scan a single module.
-
-    Args:
-        request_factory: The factory for creating requests
-        module: The prompt module to scan
-        fuzzer_state: State tracking object for the fuzzer
-        processed_prompts: Number of prompts processed so far
-        total_prompts: Total number of prompts to process
-        max_budget: Maximum token budget
-        total_tokens: Current token count
-        optimize: Whether to use optimization
-        stop_event: Event to stop scanning
-
-    Yields:
-        ScanResult objects as the scan progresses
-    """
-    tokens = 0
-    module_failures = 0
-    module_prompts = 0
-    failure_rates = []
-    should_stop = False
-
-    # Initialize optimizer if optimization is enabled
-    optimizer = (
-        Optimizer(
-            [Real(0, 1)], base_estimator="GP", n_initial_points=INITIAL_OPTIMIZER_POINTS
-        )
-        if optimize
-        else None
-    )
-
-    module_size = 0 if module.lazy else len(module.prompts)
-    logger.info(f"Scanning {module.dataset_name} {module_size}")
-    yield ScanResult(
-        module=module.dataset_name,
-        tokens=0,
-        cost=0,
-        progress=0,
-        failureRate=0,
-        prompt="",
-        latency=0,
-        model="",
-    ).model_dump_json()
-
-    async for prompt in generate_prompts(module.prompts):
-        if stop_event and stop_event.is_set():
-            stop_event.clear()
-            logger.info("Scan stopped by user.")
-            yield ScanResult.status_msg("Scan stopped by user.")
-            return
-
-        processed_prompts += 1
-        module_prompts += 1
-
-        # Calculate progress based on total processed prompts
-        progress = 100 * processed_prompts / total_prompts if total_prompts else 0
-        progress = progress % 100
-
-        total_tokens -= tokens
-        start = time.time()
-
-        tokens, failed = await process_prompt(
-            request_factory,
-            prompt,
-            tokens,
-            module.dataset_name,
-            fuzzer_state=fuzzer_state,
-        )
-
-        end = time.time()
-        total_tokens += tokens
-
-        if failed:
-            module_failures += 1
-
-        failure_rate = module_failures / max(module_prompts, 1)
-        failure_rates.append(failure_rate)
-        cost = calculate_cost(tokens)
-
-        response_text = fuzzer_state.get_last_output(prompt) or ""
-
-        yield ScanResult(
-            module=module.dataset_name,
-            tokens=round(tokens / 1000, 1),
-            cost=cost,
-            progress=round(progress, 2),
-            failureRate=round(failure_rate * 100, 2),
-            prompt=prompt[:MAX_PROMPT_LENGTH],
-            latency=end - start,
-            model=response_text,
-        ).model_dump_json()
-
-        # Optimization logic
-        if optimize and optimizer and len(failure_rates) >= MIN_FAILURE_SAMPLES:
-            next_point = optimizer.ask()
-            optimizer.tell(next_point, -failure_rate)
-            best_failure_rate = -optimizer.get_result().fun
-            if best_failure_rate > FAILURE_RATE_THRESHOLD:
-                yield ScanResult.status_msg(
-                    f"High failure rate detected ({best_failure_rate:.2%}). Stopping this module..."
-                )
-                should_stop = True
-                break
-
-        # Budget check
-        if total_tokens > max_budget:
-            logger.info(
-                f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
-            )
-            yield ScanResult.status_msg(
-                f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
-            )
-            should_stop = True
-            break
-
-        if should_stop:
-            break
-
-    return
-
-
-async def with_error_handling(agen):
-    """
-    Wraps an asynchronous generator with error handling.
-
-    This function iterates over an asynchronous generator, yielding its values.
-    If an exception occurs, it logs the error and yields a failure message.
-    Finally, it ensures that a completion message is always yielded.
-
-    Args:
-        agen: An asynchronous generator that produces scan results.
-
-    Yields:
-        ScanResult: Either a successful result, an error message if an
-        exception occurs, or a completion message at the end.
-    """
-    try:
-        async for t in agen:
-            yield t
-    except Exception as e:
-        logger.exception("Scan failed")
-        yield ScanResult.status_msg(f"Scan failed: {str(e)}")
-    finally:
-        yield ScanResult.status_msg("Scan completed.")
-
-
 async def perform_single_shot_scan(
     request_factory,
     max_budget: int,
     datasets: list[dict[str, str]] = [],
     tools_inbox=None,
-    optimize: bool = False,
-    stop_event: asyncio.Event | None = None,
+    optimize=False,
+    stop_event: asyncio.Event = None,
     secrets: dict[str, str] = {},
 ) -> AsyncGenerator[str, None]:
-    """
-    Perform a standard security scan using a given request factory.
-
-    This function processes security scan prompts from selected datasets while
-    respecting a predefined token budget. It supports optimization, failure tracking,
-    and early stopping based on budget constraints or user intervention.
-
-    Args:
-        request_factory: A factory function that generates requests for processing prompts.
-        max_budget (int): The maximum token budget for the scan.
-        datasets (list[dict[str, str]], optional): A list of datasets containing security prompts.
-        tools_inbox: Optional additional tools for processing (default: None).
-        optimize (bool, optional): Whether to enable failure rate optimization (default: False).
-        stop_event (asyncio.Event, optional): An event to signal early termination (default: None).
-        secrets (dict[str, str], optional): A dictionary of secrets for authentication (default: {}).
-
-    Yields:
-        str: JSON-encoded scan results or status messages.
-
-    The function iterates over prompts, processes them asynchronously, and updates
-    failure statistics and token usage. If the scan exceeds the budget or failure rate is too high,
-    it stops execution. Results are saved to a CSV file upon completion.
-    """
+    """Perform a standard security scan."""
     max_budget = max_budget * BUDGET_MULTIPLIER
     selected_datasets = [m for m in datasets if m["selected"]]
-    request_factory = get_modality_adapter(request_factory)
-
-    yield ScanResult.status_msg("Loading datasets...")
-    prompt_modules = prepare_prompts(
-        dataset_names=[m["dataset_name"] for m in selected_datasets],
-        budget=max_budget,
-        tools_inbox=tools_inbox,
-        options=[m.get("opts", {}) for m in selected_datasets],
-    )
-    yield ScanResult.status_msg("Datasets loaded. Starting scan...")
-
-    fuzzer_state = FuzzerState()
-    total_prompts = sum(len(m.prompts) for m in prompt_modules if not m.lazy)
-    processed_prompts = 0
-
-    total_tokens = 0
-    for module in prompt_modules:
-        module_gen = scan_module(
-            request_factory=request_factory,
-            module=module,
-            fuzzer_state=fuzzer_state,
-            processed_prompts=processed_prompts,
-            total_prompts=total_prompts,
-            max_budget=max_budget,
-            total_tokens=total_tokens,
-            optimize=optimize,
-            stop_event=stop_event,
+    request_factory = multi_modality_spec(request_factory)
+    try:
+        yield ScanResult.status_msg("Loading datasets...")
+        prompt_modules = prepare_prompts(
+            dataset_names=[m["dataset_name"] for m in selected_datasets],
+            budget=max_budget,
+            tools_inbox=tools_inbox,
+            options=[m.get("opts", {}) for m in selected_datasets],
         )
-        try:
-            async for result in module_gen:
-                yield result
-        except Exception:
-            logger.error("Module exception")
-            continue
-        # Update processed_prompts count
-        module_size = 0 if module.lazy else len(module.prompts)
-        processed_prompts += module_size
+        yield ScanResult.status_msg("Datasets loaded. Starting scan...")
 
-    yield ScanResult.status_msg("Scan completed.")
-    fuzzer_state.export_failures("failures.csv")
+        errors = []
+        refusals = []
+        outputs = []
+        total_prompts = sum(len(m.prompts) for m in prompt_modules if not m.lazy)
+        processed_prompts = 0
+
+        optimizer = (
+            Optimizer([Real(0, 1)], base_estimator="GP", n_initial_points=25)
+            if optimize
+            else None
+        )
+        failure_rates = []
+
+        total_tokens = 0
+        tokens = 0
+        should_stop = False
+        for module in prompt_modules:
+            if should_stop:
+                break
+            tokens = 0
+            module_failures = 0
+            module_size = 0 if module.lazy else len(module.prompts)
+            logger.info(f"Scanning {module.dataset_name} {module_size}")
+            module_prompts = 0  # Reset for each module
+
+            async for prompt in generate_prompts(module.prompts):
+                if stop_event and stop_event.is_set():
+                    stop_event.clear()
+                    logger.info("Scan stopped by user.")
+                    yield ScanResult.status_msg("Scan stopped by user.")
+                    return
+
+                processed_prompts += 1
+                module_prompts += 1  # Fixed increment syntax
+                # Calculate progress based on total processed prompts
+                progress = (
+                    100 * processed_prompts / total_prompts if total_prompts else 0
+                )
+
+                total_tokens -= tokens
+                start = time.time()
+                tokens, failed = await process_prompt(
+                    request_factory,
+                    prompt,
+                    tokens,
+                    module.dataset_name,
+                    refusals,
+                    errors,
+                    outputs,
+                )
+                end = time.time()
+                total_tokens += tokens
+
+                if failed:
+                    module_failures += 1
+                failure_rate = module_failures / max(module_prompts, 1)
+                failure_rates.append(failure_rate)
+                cost = calculate_cost(tokens)
+
+                last_output = outputs[-1] if outputs else None
+                if last_output and last_output[1] == prompt:
+                    response_text = last_output[2]
+                else:
+                    response_text = ""
+
+                yield ScanResult(
+                    module=module.dataset_name,
+                    tokens=round(tokens / 1000, 1),
+                    cost=cost,
+                    progress=round(progress, 2),
+                    failureRate=round(failure_rate * 100, 2),
+                    prompt=prompt[:MAX_PROMPT_LENGTH],
+                    latency=end - start,
+                    model=response_text,
+                ).model_dump_json()
+
+                if optimize and len(failure_rates) >= 5:
+                    next_point = optimizer.ask()
+                    optimizer.tell(next_point, -failure_rate)
+                    best_failure_rate = -optimizer.get_result().fun
+                    if best_failure_rate > 0.5:
+                        yield ScanResult.status_msg(
+                            f"High failure rate detected ({best_failure_rate:.2%}). Stopping this module..."
+                        )
+                        should_stop = True
+                        break
+                if total_tokens > max_budget:
+                    logger.info(
+                        f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
+                    )
+                    yield ScanResult.status_msg(
+                        f"Scan ran out of budget and stopped. {total_tokens=} {max_budget=}"
+                    )
+                    should_stop = True
+                    break
+
+        yield ScanResult.status_msg("Scan completed.")
+
+        failure_data = errors + refusals
+        df = pd.DataFrame(
+            failure_data, columns=["module", "prompt", "status_code", "content"]
+        )
+        df.to_csv("failures.csv", index=False)
+
+    except Exception as e:
+        logger.exception("Scan failed")
+        yield ScanResult.status_msg(f"Scan failed: {str(e)}")
+    finally:
+        yield ScanResult.status_msg("Scan completed.")
 
 
 async def perform_many_shot_scan(
@@ -419,180 +247,148 @@ async def perform_many_shot_scan(
     datasets: list[dict[str, str]] = [],
     probe_datasets: list[dict[str, str]] = [],
     tools_inbox=None,
-    optimize: bool = False,
-    stop_event: asyncio.Event | None = None,
+    optimize=False,
+    stop_event: asyncio.Event = None,
     probe_frequency: float = 0.2,
     max_ctx_length: int = 10_000,
     secrets: dict[str, str] = {},
 ) -> AsyncGenerator[str, None]:
-    """
-    Perform a multi-step security scan with probe injection.
+    """Perform a multi-step security scan with probe injection."""
+    request_factory = multi_modality_spec(request_factory)
+    try:
+        # Load main and probe datasets
+        yield ScanResult.status_msg("Loading datasets...")
+        prompt_modules = prepare_prompts(
+            dataset_names=[m["dataset_name"] for m in datasets if m["selected"]],
+            budget=max_budget,
+            tools_inbox=tools_inbox,
+        )
+        yield ScanResult.status_msg("Loading datasets for MSJ...")
+        msj_modules = msj_data.prepare_prompts(probe_datasets)
+        yield ScanResult.status_msg("Datasets loaded. Starting scan...")
 
-    This function executes a security scan while periodically injecting probe datasets
-    to test system robustness. It tracks failures, optimizes scan efficiency,
-    and ensures adherence to a predefined token budget.
+        errors = []
+        refusals = []
+        outputs = []
+        total_prompts = sum(len(m.prompts) for m in prompt_modules if not m.lazy)
+        processed_prompts = 0
 
-    Args:
-        request_factory: A factory function that generates requests for processing prompts.
-        max_budget (int): The maximum token budget for the scan.
-        datasets (list[dict[str, str]], optional): The main datasets for scanning.
-        probe_datasets (list[dict[str, str]], optional): Additional datasets for probe injection.
-        tools_inbox: Optional tools for additional processing (default: None).
-        optimize (bool, optional): Whether to enable failure rate optimization (default: False).
-        stop_event (asyncio.Event, optional): An event to signal early termination (default: None).
-        probe_frequency (float, optional): The probability of probe injection (default: 0.2).
-        max_ctx_length (int, optional): The maximum context length before resetting (default: 10,000 tokens).
-        secrets (dict[str, str], optional): A dictionary of secrets for authentication (default: {}).
+        optimizer = (
+            Optimizer([Real(0, 1)], base_estimator="GP", n_initial_points=25)
+            if optimize
+            else None
+        )
+        failure_rates = []
 
-    Yields:
-        str: JSON-encoded scan results or status messages.
+        for module in prompt_modules:
+            module_failures = 0
+            module_size = 0 if module.lazy else len(module.prompts)
+            logger.info(f"Scanning {module.dataset_name} {module_size}")
 
-    This function iterates over prompts, injects probe prompts at random intervals,
-    processes them asynchronously, and tracks failure rates. If failure rates exceed a threshold
-    or budget is exhausted, the scan is stopped early. Results are saved to a CSV file upon completion.
-    """
-    request_factory = get_modality_adapter(request_factory)
-    # Load main and probe datasets
-    yield ScanResult.status_msg("Loading datasets...")
-    prompt_modules = prepare_prompts(
-        dataset_names=[m["dataset_name"] for m in datasets if m["selected"]],
-        budget=max_budget,
-        tools_inbox=tools_inbox,
-    )
-    yield ScanResult.status_msg("Loading datasets for MSJ...")
-    msj_modules = msj_data.prepare_prompts(probe_datasets)
-    yield ScanResult.status_msg("Datasets loaded. Starting scan...")
-
-    fuzzer_state = FuzzerState()
-    total_prompts = sum(len(m.prompts) for m in prompt_modules if not m.lazy)
-    processed_prompts = 0
-
-    failure_rates = []
-
-    for module in prompt_modules:
-        module_failures = 0
-        module_size = 0 if module.lazy else len(module.prompts)
-        logger.info(f"Scanning {module.dataset_name} {module_size}")
-
-        async for prompt in generate_prompts(module.prompts):
-            if stop_event and stop_event.is_set():
-                stop_event.clear()
-                logger.info("Scan stopped by user.")
-                yield ScanResult.status_msg("Scan stopped by user.")
-                return
-
-            tokens = 0
-            processed_prompts += 1
-            progress = 100 * processed_prompts / total_prompts if total_prompts else 0
-            progress = progress % 100
-
-            full_prompt = ""
-            msj_module = random.choice(msj_modules)
-
-            prompt_tokens = len(full_prompt.split())
-            tokens += prompt_tokens
-
-            injected = False
-            for _ in range(20):
-                if injected:
-                    break
-
-                m_prompt = random.choice(msj_module.prompts)
-                full_prompt += "\n" + m_prompt
-                if tokens > max_ctx_length:
-                    full_prompt = "\n" + prompt
-                    injected = True
-
-                tokens, failed = await process_prompt(
-                    request_factory,
-                    full_prompt,
-                    tokens,
-                    module.dataset_name,
-                    fuzzer_state=fuzzer_state,
+            async for prompt in generate_prompts(module.prompts):
+                if stop_event and stop_event.is_set():
+                    stop_event.clear()
+                    logger.info("Scan stopped by user.")
+                    yield ScanResult.status_msg("Scan stopped by user.")
+                    return
+                tokens = 0
+                processed_prompts += 1
+                progress = (
+                    100 * processed_prompts / total_prompts if total_prompts else 0
                 )
-                if failed:
-                    module_failures += 1
-                    break
-                if injected:
-                    break
 
-            failure_rate = module_failures / max(processed_prompts, 1)
-            failure_rates.append(failure_rate)
-            cost = calculate_cost(tokens)
+                full_prompt = ""
+                msj_module = random.choice(msj_modules)
 
-            yield ScanResult(
-                module=module.dataset_name,
-                tokens=round(tokens / 1000, 1),
-                cost=cost,
-                progress=round(progress, 2),
-                failureRate=round(failure_rate * 100, 2),
-                prompt=prompt[:MAX_PROMPT_LENGTH],
-            ).model_dump_json()
+                prompt_tokens = len(full_prompt.split())
+                tokens += prompt_tokens
 
-            if optimize and len(failure_rates) >= MIN_FAILURE_SAMPLES:
-                yield ScanResult.status_msg(
-                    f"High failure rate detected ({failure_rate:.2%}). Stopping this module..."
-                )
-                break
+                injected = False
+                for _ in range(20):
+                    if injected:
+                        break
 
-    yield ScanResult.status_msg("Scan completed.")
-    fuzzer_state.export_failures("failures.csv")
+                    m_prompt = random.choice(msj_module.prompts)
+                    full_prompt += "\n" + m_prompt
+                    if tokens > max_ctx_length:
+                        full_prompt = "\n" + prompt
+                        injected = True
+
+                    tokens, failed = await process_prompt(
+                        request_factory,
+                        full_prompt,
+                        tokens,
+                        module.dataset_name,
+                        refusals,
+                        errors,
+                        outputs,
+                    )
+                    if failed:
+                        module_failures += 1
+                        break
+                    if injected:
+                        break
+
+                failure_rate = module_failures / max(processed_prompts, 1)
+                failure_rates.append(failure_rate)
+                cost = calculate_cost(tokens)
+
+                yield ScanResult(
+                    module=module.dataset_name,
+                    tokens=round(tokens / 1000, 1),
+                    cost=cost,
+                    progress=round(progress, 2),
+                    failureRate=round(failure_rate * 100, 2),
+                    prompt=prompt[:MAX_PROMPT_LENGTH],
+                ).model_dump_json()
+
+                if optimize and len(failure_rates) >= 5:
+                    next_point = optimizer.ask()
+                    optimizer.tell(next_point, -failure_rate)
+                    best_failure_rate = -optimizer.get_result().fun
+                    if best_failure_rate > 0.5:
+                        yield ScanResult.status_msg(
+                            f"High failure rate detected ({best_failure_rate:.2%}). Stopping this module..."
+                        )
+                        break
+
+        yield ScanResult.status_msg("Scan completed.")
+
+        df = pd.DataFrame(
+            errors + refusals, columns=["module", "prompt", "status_code", "content"]
+        )
+        df.to_csv("failures.csv", index=False)
+
+    except Exception as e:
+        logger.exception("Scan failed")
+        yield ScanResult.status_msg(f"Scan failed: {str(e)}")
+        raise e
 
 
 def scan_router(
     request_factory,
     scan_parameters: Scan,
     tools_inbox=None,
-    stop_event: asyncio.Event | None = None,
+    stop_event: asyncio.Event = None,
 ):
-    """
-    Route scan requests to the appropriate scanning function.
-
-    This function determines whether to perform a multi-step or single-shot
-    security scan based on the provided scan parameters.
-
-    Args:
-        request_factory: A factory function to generate requests for processing prompts.
-        scan_parameters (Scan): An object containing the parameters for the scan, including:
-            - enableMultiStepAttack (bool): Whether to perform a multi-step scan.
-            - maxBudget (int): The maximum token budget for the scan.
-            - datasets (list[dict[str, str]]): The datasets to scan.
-            - probe_datasets (list[dict[str, str]], optional): Datasets for probe injection (multi-step only).
-            - optimize (bool): Whether to enable optimization.
-            - secrets (dict[str, str], optional): A dictionary of secrets for authentication.
-        tools_inbox: Optional tools for additional processing (default: None).
-        stop_event (asyncio.Event, optional): An event to signal early termination (default: None).
-
-    Returns:
-        A function wrapped with `with_error_handling`, which executes either:
-        - `perform_many_shot_scan` for multi-step scanning.
-        - `perform_single_shot_scan` for single-shot scanning.
-
-    The function ensures that the appropriate scanning method is chosen based on
-    the `enableMultiStepAttack` flag in `scan_parameters`.
-    """
     if scan_parameters.enableMultiStepAttack:
-        return with_error_handling(
-            perform_many_shot_scan(
-                request_factory=request_factory,
-                max_budget=scan_parameters.maxBudget,
-                datasets=scan_parameters.datasets,
-                probe_datasets=scan_parameters.probe_datasets,
-                tools_inbox=tools_inbox,
-                optimize=scan_parameters.optimize,
-                stop_event=stop_event,
-                secrets=scan_parameters.secrets,
-            )
+        return perform_many_shot_scan(
+            request_factory=request_factory,
+            max_budget=scan_parameters.maxBudget,
+            datasets=scan_parameters.datasets,
+            probe_datasets=scan_parameters.probe_datasets,
+            tools_inbox=tools_inbox,
+            optimize=scan_parameters.optimize,
+            stop_event=stop_event,
+            secrets=scan_parameters.secrets,
         )
     else:
-        return with_error_handling(
-            perform_single_shot_scan(
-                request_factory=request_factory,
-                max_budget=scan_parameters.maxBudget,
-                datasets=scan_parameters.datasets,
-                tools_inbox=tools_inbox,
-                optimize=scan_parameters.optimize,
-                stop_event=stop_event,
-                secrets=scan_parameters.secrets,
-            )
+        return perform_single_shot_scan(
+            request_factory=request_factory,
+            max_budget=scan_parameters.maxBudget,
+            datasets=scan_parameters.datasets,
+            tools_inbox=tools_inbox,
+            optimize=scan_parameters.optimize,
+            stop_event=stop_event,
+            secrets=scan_parameters.secrets,
         )

agentic_security/probe_actor/operator.py

@@ -1,4 +1,5 @@
 import asyncio
+import logging
 from typing import Any
 
 import httpx
@@ -6,10 +7,13 @@ from pydantic import BaseModel, Field
 from pydantic_ai import Agent, RunContext
 
 from agentic_security.http_spec import LLMSpec
-from agentic_security.logutils import logger
 
 LLM_SPECS = []
 
+# Configure logging
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
 
 class AgentSpecification(BaseModel):
     name: str | None = Field(None, description="Name of the LLM/agent")

agentic_security/probe_actor/refusal.py

@@ -1,6 +1,6 @@
 from abc import ABC, abstractmethod
 
-from agentic_security.refusal_classifier.model import RefusalClassifier
+from agentic_security.refusal_classifier import RefusalClassifier
 
 classifier = RefusalClassifier()
 classifier.load_model()

agentic_security/probe_actor/state.py

@@ -1,47 +0,0 @@
-import pandas as pd
-
-
-class FuzzerState:
-    """Container for tracking scan results"""
-
-    def __init__(self):
-        self.errors = []
-        self.refusals = []
-        self.outputs = []
-
-    def add_error(
-        self,
-        module_name: str,
-        prompt: str,
-        status_code: int | str,
-        error_msg: str,
-    ):
-        """Add an error to the state"""
-        self.errors.append((module_name, prompt, status_code, error_msg))
-
-    def add_refusal(
-        self, module_name: str, prompt: str, status_code: int, response_text: str
-    ):
-        """Add a refusal to the state"""
-        self.refusals.append((module_name, prompt, status_code, response_text))
-
-    def add_output(
-        self, module_name: str, prompt: str, response_text: str, refused: bool
-    ):
-        """Add an output to the state"""
-        self.outputs.append((module_name, prompt, response_text, refused))
-
-    def get_last_output(self, prompt: str) -> str | None:
-        """Get the last output for a given prompt"""
-        for output in reversed(self.outputs):
-            if output[1] == prompt:
-                return output[2]
-        return None
-
-    def export_failures(self, filename: str = "failures.csv"):
-        """Export failures to a CSV file"""
-        failure_data = self.errors + self.refusals
-        df = pd.DataFrame(
-            failure_data, columns=["module", "prompt", "status_code", "content"]
-        )
-        df.to_csv(filename, index=False)

agentic_security/probe_actor/test_fuzzer.py

@@ -7,7 +7,6 @@ import pytest
 
 from agentic_security.primitives import Scan
 from agentic_security.probe_actor.fuzzer import (
-    FuzzerState,
     generate_prompts,
     perform_many_shot_scan,
     perform_single_shot_scan,
@@ -208,7 +207,9 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            fuzzer_state=FuzzerState(),
+            refusals=[],
+            errors=[],
+            outputs=[],
         )
 
         self.assertEqual(tokens, 3)  # Tokens from "Valid response text"
@@ -225,17 +226,20 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             )
         )
 
-        fuzzer_state = FuzzerState()
+        refusals = []
+        outputs = []
         tokens, refusal = await process_prompt(
             request_factory=mock_request_factory,
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            fuzzer_state=fuzzer_state,
+            refusals=refusals,
+            errors=[],
+            outputs=outputs,
         )
 
         self.assertEqual(tokens, 3)  # Tokens from "Response indicating refusal"
-        # self.assertFalse(fuzzer_state.refusals)
+        self.assertFalse(refusal)
 
     async def test_http_error_response(self):
         mock_request_factory = Mock()
@@ -248,13 +252,15 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             )
         )
 
-        fuzzer_state = FuzzerState()
+        refusals = []
         await process_prompt(
             request_factory=mock_request_factory,
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            fuzzer_state=fuzzer_state,
+            refusals=refusals,
+            errors=[],
+            outputs=[],
         )
 
     async def test_request_error(self):
@@ -263,14 +269,18 @@ class TestProcessPrompt(unittest.IsolatedAsyncioTestCase):
             side_effect=httpx.RequestError("Connection error")
         )
 
-        fuzzer_state = FuzzerState()
+        errors = []
         tokens, refusal = await process_prompt(
             request_factory=mock_request_factory,
             prompt="test prompt",
             tokens=0,
             module_name="module_a",
-            fuzzer_state=fuzzer_state,
+            refusals=[],
+            errors=errors,
+            outputs=[],
         )
 
         self.assertEqual(tokens, 0)
         self.assertTrue(refusal)
+        self.assertEqual(len(errors), 1)
+        self.assertIn("Connection error", errors[0][3])

agentic_security/probe_data/__init__.py

@@ -1,4 +1,4 @@
-from .data import load_local_csv, load_local_csv_files
+from .data import load_local_csv
 
 REGISTRY_V0 = [
     {
@@ -484,18 +484,3 @@ REGISTRY = REGISTRY_V0 + [
         "modality": "text",
     },
 ]
-
-for ds in load_local_csv_files():
-    REGISTRY.append(
-        {
-            "dataset_name": ds.dataset_name,
-            "num_prompts": len(ds.prompts),
-            "tokens": ds.prompts,
-            "approx_cost": 0.0,
-            "is_active": True,
-            "source": f"Local file dataset: {ds.metadata['src']}",
-            "selected": False,
-            "url": "",
-            "modality": "text",
-        }
-    )

agentic_security/probe_data/audio_generator.py

@@ -1,5 +1,4 @@
 import base64
-import logging
 import os
 import platform
 import subprocess
@@ -8,16 +7,6 @@ import uuid
 import httpx
 from cache_to_disk import cache_to_disk
 
-# Configure logging
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-class AudioGenerationError(Exception):
-    """Custom exception for errors during audio generation."""
-
-    pass
-
 
 def encode(content: bytes) -> str:
     encoded_content = base64.b64encode(content).decode("utf-8")
@@ -52,25 +41,12 @@ def generate_audio_mac_wav(prompt: str) -> bytes:
         # Read the WAV file into memory
         with open(temp_wav_path, "rb") as f:
             audio_bytes = f.read()
-
-    except subprocess.CalledProcessError as e:
-        logger.error(f"Subprocess error: {e}")
-        raise AudioGenerationError("Failed to generate or convert audio.") from e
-    except FileNotFoundError as e:
-        logger.error(f"File not found: {e}")
-        raise AudioGenerationError("Required file not found.") from e
-    except Exception as e:
-        logger.exception("Unexpected error occurred.")
-        raise AudioGenerationError(
-            "An unexpected error occurred during audio generation."
-        ) from e
     finally:
-        for path in (temp_aiff_path, temp_wav_path):
-            try:
-                if os.path.exists(path):
-                    os.remove(path)
-            except Exception as e:
-                logger.warning(f"Failed to delete temporary file {path}: {e}")
+        # Clean up the temporary files
+        if os.path.exists(temp_aiff_path):
+            os.remove(temp_aiff_path)
+        if os.path.exists(temp_wav_path):
+            os.remove(temp_wav_path)
 
     # Return the audio bytes
     return audio_bytes

agentic_security/probe_data/data.py

@@ -1,16 +1,14 @@
 import io
 import os
 import random
-from collections.abc import Callable, Iterator
-from functools import partial
-from typing import Any, TypeVar
+from functools import lru_cache
 
 import httpx
 import pandas as pd
 from cache_to_disk import cache_to_disk
 from datasets import load_dataset
+from loguru import logger
 
-from agentic_security.logutils import logger
 from agentic_security.probe_data import stenography_fn
 from agentic_security.probe_data.models import ProbeDataset
 from agentic_security.probe_data.modules import (
@@ -21,279 +19,281 @@ from agentic_security.probe_data.modules import (
     rl_model,
 )
 
-# Type aliases for clarity
-T = TypeVar("T")
-FilterFn = Callable[[pd.Series], bool]
-ColumnMappings = dict[str, str]
-DatasetLoader = Callable[[], ProbeDataset]
-TransformFn = Callable[[str], str]
 
+@cache_to_disk()
+def load_dataset_general(
+    dataset_name,
+    dataset_split="train",
+    column_mappings=None,
+    filter_fn=None,
+    custom_url=None,
+    additional_metadata=None,
+):
+    """
+    Generalized function to load datasets with flexible configurations.
 
-# Core data loading utilities
-def fetch_csv_content(url: str) -> str:
-    """Fetch CSV content from a URL."""
-    response = httpx.get(url)
-    response.raise_for_status()  # Raise exception for bad responses
-    return response.content.decode("utf-8")
+    :param dataset_name: Name of the dataset or URL for custom CSVs
+    :param dataset_split: Split to load from the dataset (e.g., "train")
+    :param column_mappings: Dictionary mapping dataset columns to expected keys, e.g., {'prompt': 'query'}
+    :param filter_fn: A filtering function that takes a row and returns True/False
+    :param custom_url: URL for custom CSV datasets
+    :param additional_metadata: Additional metadata to include in the ProbeDataset
+    :return: A ProbeDataset object with the processed data
+    """
+    if custom_url:
+        logger.info(f"Loading custom CSV dataset from {custom_url}")
+        r = httpx.get(custom_url)
+        content = r.content
+        df = pd.read_csv(io.StringIO(content.decode("utf-8")))
+    else:
+        logger.info(f"Loading dataset {dataset_name} from Hugging Face datasets")
+        dataset = load_dataset(dataset_name)
+        df = pd.DataFrame(dataset[dataset_split])
 
+    # Apply column mappings if provided
+    if column_mappings:
+        df.rename(columns=column_mappings, inplace=True)
 
-def load_df_from_source(source: str, is_url: bool = False) -> pd.DataFrame:
-    """Load DataFrame from either URL or Hugging Face dataset."""
-    if is_url:
-        logger.info(f"Loading custom CSV from {source}")
-        return pd.read_csv(io.StringIO(fetch_csv_content(source)))
-    logger.info(f"Loading dataset {source} from Hugging Face")
-    return pd.DataFrame(load_dataset(source)["train"])
-
-
-def transform_df(
-    df: pd.DataFrame,
-    mappings: ColumnMappings | None = None,
-    filter_fn: FilterFn | None = None,
-) -> pd.DataFrame:
-    """Apply transformations to DataFrame."""
-    result = df.copy()
-    if mappings:
-        result.rename(columns=mappings, inplace=True)
+    # Filter rows if filter_fn is provided
     if filter_fn:
-        result = result[result.apply(filter_fn, axis=1)]
-    return result
+        df = df[df.apply(filter_fn, axis=1)]
 
+    # Extract prompts
+    prompts = df[column_mappings.get("prompt", "prompt")].tolist()
 
-def create_probe_dataset(
-    name: str, prompts: list[str], metadata: dict[str, Any] | None = None
-) -> ProbeDataset:
-    """Create a ProbeDataset from prompts."""
-    metadata = metadata or {}
     return ProbeDataset(
-        dataset_name=name,
-        metadata=metadata,
+        dataset_name=dataset_name,
+        metadata=additional_metadata or {},
         prompts=prompts,
-        tokens=sum(len(str(s).split()) for s in prompts),
+        tokens=count_words_in_list(prompts),
         approx_cost=0.0,
     )
 
 
-# Generalized dataset loader
+def count_words_in_list(str_list):
+    """Calculate the total number of words in a given list of strings.
+
+    :param str_list: List of strings
+    :return: Total number of words across all strings in the list
+    """
+    total_words = sum(len(str(s).split()) for s in str_list)
+    return total_words
+
+
 @cache_to_disk()
-def load_dataset_generic(
-    name: str,
-    mappings: ColumnMappings | None = None,
-    filter_fn: FilterFn | None = None,
-    url: str | None = None,
-    metadata: dict[str, Any] | None = None,
-) -> ProbeDataset:
-    """Load and process a dataset with flexible configuration."""
-    try:
-        df = load_df_from_source(url or name, is_url=bool(url))
-        transformed_df = transform_df(df, mappings, filter_fn)
-
-        # Determine which column to use as the prompt source
-        prompt_col = None
-        if mappings and "prompt" in mappings:
-            prompt_col = mappings["prompt"]
-        elif "prompt" in transformed_df.columns:
-            prompt_col = "prompt"
-        else:
-            # Try to find a suitable text column
-            text_columns = [
-                col
-                for col in transformed_df.columns
-                if any(
-                    keyword in col.lower()
-                    for keyword in ["prompt", "text", "query", "question"]
-                )
-            ]
-            if text_columns:
-                prompt_col = text_columns[0]
-                logger.info(f"Using column '{prompt_col}' as prompt source")
-            else:
-                logger.error(f"No suitable prompt column found in dataset {name}")
-                return create_probe_dataset(name, [], metadata)
-
-        # Extract prompts and filter out empty ones
-        prompts = [
-            p
-            for p in transformed_df[prompt_col].tolist()
-            if p and isinstance(p, (str, int, float))
-        ]
-        return create_probe_dataset(name, prompts, metadata)
-    except Exception as e:
-        logger.error(f"Error loading dataset {name}: {e}")
-        return create_probe_dataset(name, [], {"error": str(e)})
-
-
-# Dataset-specific configurations
-DATASET_CONFIGS = {
-    "ShawnMenz/DAN_jailbreak": {
-        "mappings": {"prompt": "prompt"},
-        "filter_fn": lambda row: row["jailbreak"],
-    },
-    "deepset/prompt-injections": {
-        "mappings": {"prompt": "text"},
-        "filter_fn": lambda row: row["label"],
-    },
-    "notrichardren/refuse-to-answer-prompts": {
-        "mappings": {"prompt": "claim"},
-        "filter_fn": lambda row: row["label"],
-    },
-    "rubend18/ChatGPT-Jailbreak-Prompts": {
-        "mappings": {"prompt": "Prompt"},
-        "filter_fn": lambda row: row["label"],
-    },
-    "markush1/LLM-Jailbreak-Classifier": {
-        "mappings": {"prompt": "prompt"},
-        "filter_fn": lambda row: row["classification"] == "jailbreak",
-    },
-    "ShawnMenz/jailbreak_sft_rm_ds": {
-        "url": "hf://datasets/ShawnMenz/jailbreak_sft_rm_ds/jailbreak_sft_rm_ds.csv",
-        "mappings": {"prompt": "prompt"},
-        "filter_fn": lambda row: row["jailbreak"] == "jailbreak",
-    },
-    "verazuo/jailbreak_llms/2023_05_07": {
-        "url": "https://raw.githubusercontent.com/verazuo/jailbreak_llms/main/data/prompts/jailbreak_prompts_2023_05_07.csv",
-        "mappings": {"prompt": "prompt"},
-        "filter_fn": lambda row: bool(row["jailbreak"]),
-    },
-    "verazuo/jailbreak_llms/2023_12_25": {
-        "url": "https://raw.githubusercontent.com/verazuo/jailbreak_llms/main/data/prompts/jailbreak_prompts_2023_12_25.csv",
-        "mappings": {"prompt": "prompt"},
-        "filter_fn": lambda row: bool(row["jailbreak"]),
-    },
-}
-
-# Additional generic dataset configurations
-DATASET_CONFIGS_GENERICS = {
-    "simonycl/aya-23-8B_advbench_jailbreak": {"mappings": {"prompt": "prompt"}},
-    "acmc/jailbreaks_dataset_with_perplexity_bigcode_starcoder2-3b_bigcode_starcoder2-7b": {},
-    "karanxa/dolphin-jailbreak-finetuning-dataset": {"mappings": {"prompt": "text"}},
-    "karanxa/llama-2-jailbreak-dataset": {"mappings": {"prompt": "text"}},
-    "karanxa/llama2-uncensored-jailbreak-dataset-finetuning": {
-        "mappings": {"prompt": "text"}
-    },
-    "liuyanchen1015/Llama-3.2-1B_jailbreak_responses": {
-        "mappings": {"prompt": "jailbreak_prompt_text"}
-    },
-    "liuyanchen1015/Llama-3.2-1B-Instruct_jailbreak_responses": {
-        "mappings": {"prompt": "jailbreak_prompt_text"}
-    },
-    "liuyanchen1015/Llama-3.2-1B-Instruct_jailbreak_responses_with_judgment": {
-        "mappings": {"prompt": "jailbreak_prompt_text"}
-    },
-    "jackhhao/jailbreak-classification": {"mappings": {"prompt": "prompt"}},
-    "walledai/JailbreakBench": {"mappings": {"prompt": "prompt"}},
-    "walledai/JailbreakHub": {"mappings": {"prompt": "prompt"}},
-    "Granther/evil-jailbreak": {"mappings": {"prompt": "text"}},
-    "sevdeawesome/jailbreak_success": {"mappings": {"prompt": "jailbreak_prompt_text"}},
-    "IDA-SERICS/Disaster-tweet-jailbreaking": {"mappings": {"prompt": "prompt_attack"}},
-    "GeorgeDaDude/Jailbreak_Complete_DS_labeled": {"mappings": {"prompt": "question"}},
-    "dayone3nder/jailbreak_prompt_JBB_sft_trainset": {"mappings": {"prompt": "prompt"}},
-    "dayone3nder/general_safe_mix_jailbreak_prompt_JBB_trainset": {
-        "mappings": {"prompt": "prompt"}
-    },
-}
-
-
-# Dataset factory
-def create_dataset_loader(name: str, config: dict[str, Any]) -> DatasetLoader:
-    """Create a dataset loader from configuration."""
-    return partial(
-        load_dataset_generic,
-        name=name,
-        mappings=config.get("mappings"),
-        filter_fn=config.get("filter_fn"),
-        url=config.get("url"),
-        metadata={"source": name, "config": str(config)},
-    )
-
-
-# Specialized loaders
-@cache_to_disk()
-def load_multi_dataset(name: str, sub_datasets: list[str]) -> ProbeDataset:
-    """Load and combine multiple sub-datasets."""
-    prompts = []
-    for sub in sub_datasets:
-        try:
-            dataset = load_dataset(name, sub)
-            if "query" in dataset["train"].features:
-                prompts.extend(dataset["train"]["query"])
-            else:
-                logger.warning(f"No 'query' column in {name}/{sub}")
-        except Exception as e:
-            logger.error(f"Error loading {name}/{sub}: {e}")
-
-    return create_probe_dataset(
-        f"{name}_combined", prompts, {"source": name, "sub_datasets": sub_datasets}
+def load_dataset_v1():
+    return load_dataset_general(
+        dataset_name="ShawnMenz/DAN_jailbreak",
+        column_mappings={"prompt": "prompt", "jailbreak": "jailbreak"},
+        filter_fn=lambda row: row["jailbreak"],
     )
 
 
 @cache_to_disk()
-def load_jailbreak_v28k() -> ProbeDataset:
-    """Load JailBreakV-28K dataset."""
-    try:
-        df = pd.read_csv(
-            "hf://datasets/JailbreakV-28K/JailBreakV-28k/JailBreakV_28K.csv"
-        )
-        prompts = df["jailbreak_query"].tolist()
-        return create_probe_dataset(
-            "JailbreakV-28K/JailBreakV-28k",
-            prompts,
-            {"source": "JailbreakV-28K/JailBreakV-28k"},
-        )
-    except Exception as e:
-        logger.error(f"Error loading JailbreakV-28K: {e}")
-        return create_probe_dataset("JailbreakV-28K/JailBreakV-28k", [])
+def load_dataset_v2():
+    return load_dataset_general(
+        dataset_name="deepset/prompt-injections",
+        column_mappings={"prompt": "text", "jailbreak": "label"},
+        filter_fn=lambda row: row["label"],
+    )
 
 
-@cache_to_disk(1)
-def file_dataset(file) -> list[str]:
-    prompts = []
-    try:
-        df = pd.read_csv(os.path.join("./datasets", file), encoding_errors="ignore")
-        if "prompt" in df.columns:
-            prompts = df["prompt"].tolist()
-        else:
-            logger.warning(f"File {file} lacks a suitable prompt column")
-    except Exception as e:
-        logger.error(f"Error reading {file}: {e}")
-    return prompts
+@cache_to_disk()
+def load_dataset_v4():
+    return load_dataset_general(
+        dataset_name="notrichardren/refuse-to-answer-prompts",
+        column_mappings={"prompt": "claim", "jailbreak": "label"},
+        filter_fn=lambda row: row["label"],
+    )
 
 
-def load_local_csv() -> ProbeDataset:
-    """Load prompts from local CSV files."""
-    os.makedirs("./datasets", exist_ok=True)
-    csv_files = [f for f in os.listdir("./datasets") if f.endswith(".csv")]
-    logger.info(f"Found {len(csv_files)} CSV files: {csv_files}")
-
-    prompts = []
-    for file in csv_files:
-        prompts.extend(file_dataset(file))
-    return create_probe_dataset("Local CSV", prompts, {"src": str(csv_files)})
+@cache_to_disk()
+def load_dataset_v3():
+    return load_dataset_general(
+        dataset_name="rubend18/ChatGPT-Jailbreak-Prompts",
+        column_mappings={"prompt": "Prompt"},
+        filter_fn=lambda row: row["label"],
+    )
 
 
-def load_csv(file: str) -> ProbeDataset:
-    """Load prompts from local CSV files."""
-    prompts = file_dataset(file)
-    return create_probe_dataset(f"fs://{file}", prompts, {"src": str(file)})
+@cache_to_disk()
+def load_dataset_v6():
+    dataset = load_dataset("markush1/LLM-Jailbreak-Classifier")
+    bad_prompts = [
+        p
+        for p, c in zip(dataset["train"]["prompt"], dataset["train"]["classification"])
+        if c == "jailbreak"
+    ]
+    return ProbeDataset(
+        dataset_name="markush1/LLM-Jailbreak-Classifier",
+        metadata={},
+        prompts=bad_prompts,
+        tokens=count_words_in_list(bad_prompts),
+        approx_cost=0.0,
+    )
 
 
-def load_local_csv_files() -> list[ProbeDataset]:
-    """Load prompts from local CSV files and return a list of ProbeDataset objects."""
-    csv_files = [f for f in os.listdir("./datasets") if f.endswith(".csv")]
-    logger.info(f"Found {len(csv_files)} CSV files: {csv_files}")
-
-    datasets = []
-
-    for file in csv_files:
-        datasets.append(create_probe_dataset(file, file_dataset(file), {"src": file}))
-    return datasets
+@cache_to_disk()
+def load_dataset_v7():
+    splits = {
+        "mini_JailBreakV_28K": "JailBreakV_28K/mini_JailBreakV_28K.csv",
+        "JailBreakV_28K": "JailBreakV_28K/JailBreakV_28K.csv",
+    }
+    df = pd.read_csv(
+        "hf://datasets/JailbreakV-28K/JailBreakV-28k/" + splits["JailBreakV_28K"]
+    )
+    bad_prompts = df["jailbreak_query"].tolist()
+    print(df.shape)
+    return ProbeDataset(
+        dataset_name="JailbreakV-28K/JailBreakV-28k",
+        metadata={},
+        prompts=bad_prompts,
+        tokens=count_words_in_list(bad_prompts),
+        approx_cost=0.0,
+    )
 
 
-# Stenography transformer
-class StenographyTransformer:
-    """Apply stenography transformations to datasets."""
+@cache_to_disk()
+def load_dataset_v8():
+    df = pd.read_csv(
+        "hf://datasets/ShawnMenz/jailbreak_sft_rm_ds/jailbreak_sft_rm_ds.csv",
+        names=["jailbreak", "prompt"],
+    )
+    filtered = df[df["jailbreak"] == "jailbreak"]["prompt"].tolist()
+    return ProbeDataset(
+        dataset_name="JailbreakV-28K/JailBreakV-28k",
+        metadata={},
+        prompts=filtered,
+        tokens=count_words_in_list(filtered),
+        approx_cost=0.0,
+    )
 
-    TRANSFORMATIONS = {
+
+@cache_to_disk()
+def load_dataset_v5():
+    ds = []
+    for c in [
+        "AdvBench",
+        "ForbiddenQuestion",
+        "MJP",
+        "MaliciousInstruct",
+        "QuestionList",
+    ]:
+        dataset = load_dataset("Lemhf14/EasyJailbreak_Datasets", c)
+        bad_prompts = dataset["train"]["query"]
+        ds.extend(bad_prompts)
+
+    return ProbeDataset(
+        dataset_name="Lemhf14/EasyJailbreak_Datasets",
+        metadata={},
+        prompts=ds,
+        tokens=count_words_in_list(ds),
+        approx_cost=0.0,
+    )
+
+
+@cache_to_disk()
+def load_generic_csv(url, name, column="prompt", predicator=None):
+    r = httpx.get(url)
+    content = r.content
+    df = pd.read_csv(io.StringIO(content.decode("utf-8")))
+    logger.info(f"Loaded {len(df)} prompts from {url}")
+    filtered_prompts = df[df.apply(predicator, axis=1)][column].tolist()
+    return ProbeDataset(
+        dataset_name=name,
+        metadata={},
+        prompts=filtered_prompts,
+        tokens=count_words_in_list(filtered_prompts),
+        approx_cost=0.0,
+    )
+
+
+def prepare_prompts(dataset_names, budget, tools_inbox=None, options=[]):
+    # ## Datasets used and cleaned:
+    # markush1/LLM-Jailbreak-Classifier
+    # 1. Open-Orca/OpenOrca
+    # 2. ShawnMenz/DAN_jailbreak
+    # 3. EddyLuo/JailBreakV_28K
+    # 4. https://raw.githubusercontent.com/verazuo/jailbreak_llms/main/data/jailbreak_prompts.csv
+
+    dataset_map = {
+        "ShawnMenz/DAN_jailbreak": load_dataset_v1,
+        "deepset/prompt-injections": load_dataset_v2,
+        "notrichardren/refuse-to-answer-prompts": load_dataset_v4,
+        "rubend18/ChatGPT-Jailbreak-Prompts": load_dataset_v3,
+        "Lemhf14/EasyJailbreak_Datasets": load_dataset_v5,
+        "markush1/LLM-Jailbreak-Classifier": load_dataset_v6,
+        "JailbreakV-28K/JailBreakV-28k": load_dataset_v7,
+        "ShawnMenz/jailbreak_sft_rm_ds": load_dataset_v8,
+        "verazuo/jailbreak_llms/2023_05_07": lambda: load_generic_csv(
+            url="https://raw.githubusercontent.com/verazuo/jailbreak_llms/main/data/prompts/jailbreak_prompts_2023_05_07.csv",
+            name="verazuo/jailbreak_llms/2023_05_07",
+            column="prompt",
+            predicator=lambda x: bool(x["jailbreak"]),
+        ),
+        "verazuo/jailbreak_llms/2023_12_25.csv": lambda: load_generic_csv(
+            url="https://raw.githubusercontent.com/verazuo/jailbreak_llms/main/data/prompts/jailbreak_prompts_2023_12_25.csv.csv",
+            name="verazuo/jailbreak_llms/2023_12_25.csv",
+            column="prompt",
+            predicator=lambda x: bool(x["jailbreak"]),
+        ),
+        "Custom CSV": load_local_csv,
+    }
+    dataset_map.update(dataset_map_generics)
+
+    group = []
+    for dataset_name in dataset_names:
+        if dataset_name in dataset_map:
+            logger.info(f"Loading {dataset_name}")
+            try:
+                group.append(dataset_map[dataset_name]())
+            except Exception as e:
+                logger.error(f"Error loading {dataset_name}: {e}")
+
+    dynamic_datasets = {
+        "AgenticBackend": lambda opts: dataset_from_iterator(
+            "AgenticBackend",
+            fine_tuned.Module(group, tools_inbox=tools_inbox, opts=opts).apply(),
+            lazy=True,
+        ),
+        "Steganography": lambda opts: Stenography(group),
+        "llm-adaptive-attacks": lambda opts: dataset_from_iterator(
+            "llm-adaptive-attacks",
+            adaptive_attacks.Module(group, tools_inbox=tools_inbox, opts=opts).apply(),
+        ),
+        "Garak": lambda opts: dataset_from_iterator(
+            "Garak",
+            garak_tool.Module(group, tools_inbox=tools_inbox, opts=opts).apply(),
+            lazy=True,
+        ),
+        "Reinforcement Learning Optimization": lambda opts: dataset_from_iterator(
+            "Reinforcement Learning Optimization",
+            rl_model.Module(group, tools_inbox=tools_inbox, opts=opts).apply(),
+            lazy=True,
+        ),
+        "InspectAI": lambda opts: dataset_from_iterator(
+            "InspectAI",
+            inspect_ai_tool.Module(group, tools_inbox=tools_inbox).apply(),
+            lazy=True,
+        ),
+        "GPT fuzzer": lambda opts: [],
+    }
+
+    dynamic_groups = []
+    options = options or [{} for _ in dataset_names]
+    for dataset_name, opts in zip(dataset_names, options):
+        if dataset_name in dynamic_datasets:
+            logger.info(f"Loading {dataset_name}")
+
+            ds = dynamic_datasets[dataset_name](opts)
+
+            for g in ds:
+                dynamic_groups.append(g)
+    return group + dynamic_groups
+
+
+class Stenography:
+    fn_library = {
         "rot5": stenography_fn.rot5,
         "rot13": stenography_fn.rot13,
         "base64": stenography_fn.base64_encode,
@@ -306,172 +306,151 @@ class StenographyTransformer:
         "zigzag_obfuscation": stenography_fn.zigzag_obfuscation,
     }
 
-    def __init__(self, datasets: list[ProbeDataset]):
-        self.datasets = datasets
+    def __init__(self, prompt_groups: [ProbeDataset]):
+        self.prompt_groups = prompt_groups
 
-    def transform(self) -> Iterator[ProbeDataset]:
-        """Generate transformed datasets."""
-        if not self.datasets:
-            logger.warning("No datasets provided to StenographyTransformer")
-            return
+    def __iter__(self):
+        return self.apply()
 
-        for dataset in self.datasets:
-            if not dataset.prompts:
-                logger.warning(
-                    f"Dataset {dataset.dataset_name} has no prompts to transform"
+    def apply(self):
+        for prompt_group in self.prompt_groups:
+            size = len(prompt_group.prompts)
+            for name, fn in self.fn_library.items():
+                logger.info(f"Applying {name} to {prompt_group.dataset_name}")
+                selected_prompts = random.sample(prompt_group.prompts, int(size * 0.25))
+                prompts = [
+                    f"Follow the instructions in {name}: {fn(p)}"
+                    for p in selected_prompts
+                ]
+                yield ProbeDataset(
+                    dataset_name=f"stenography.{name}({prompt_group.dataset_name})",
+                    metadata={},
+                    prompts=prompts,
+                    tokens=count_words_in_list(prompts),
+                    approx_cost=0.0,
                 )
-                continue
-
-            # Convert async generator (if any) to a list
-            prompts = (
-                list(dataset.prompts) if hasattr(dataset.prompts, "__iter__") else []
-            )
-            size = len(prompts)
-
-            if size == 0:
-                continue
-
-            sample_size = max(1, int(size * 0.25))  # At least 1 sample
-
-            for name, fn in self.TRANSFORMATIONS.items():
-                logger.info(f"Applying {name} to {dataset.dataset_name}")
-                sampled = random.sample(prompts, min(sample_size, size))
-                try:
-                    transformed = [
-                        f"Follow the instructions in {name}: {fn(str(p))}"
-                        for p in sampled
-                    ]
-                    transformed_dataset = create_probe_dataset(
-                        f"stenography.{name}({dataset.dataset_name})",
-                        transformed,
-                        {"source": dataset.dataset_name, "transformation": name},
-                    )
-                    yield transformed_dataset
-                except Exception as e:
-                    logger.error(
-                        f"Error applying {name} to {dataset.dataset_name}: {e}"
-                    )
 
 
-def dataset_from_iterator(
-    name: str, iterator: Iterator[str], lazy: bool = False
-) -> list[ProbeDataset]:
-    """Convert an iterator into a list of ProbeDataset objects."""
-    try:
-        prompts = list(iterator) if not lazy else iterator
-        tokens = sum(len(str(s).split()) for s in prompts) if not lazy else 0
-        dataset = ProbeDataset(
-            dataset_name=name,
-            metadata={"source": name, "lazy": lazy},
-            prompts=prompts,
-            tokens=tokens,
-            approx_cost=0.0,
-            lazy=lazy,
-        )
-        return [dataset]
-    except Exception as e:
-        logger.error(f"Error creating dataset from iterator {name}: {e}")
-        return [create_probe_dataset(name, [], {"error": str(e)})]
+@lru_cache
+def load_local_csv() -> ProbeDataset:
+    csv_files = [f for f in os.listdir(".") if f.endswith(".csv")]
+    logger.info(f"Found {len(csv_files)} CSV files")
+    logger.info(f"CSV files: {csv_files}")
+    prompt_list = []
 
-
-# Main dataset preparation
-def prepare_prompts(
-    dataset_names: list[str],
-    budget: float,
-    tools_inbox=None,
-    options: list[dict] = None,
-) -> list[ProbeDataset]:
-    """Prepare datasets based on names and options."""
-    # Base dataset loaders
-    logger.info(f"Preparing datasets: {dataset_names}")
-    dataset_loaders = {
-        **{k: create_dataset_loader(k, v) for k, v in DATASET_CONFIGS.items()},
-        **{k: create_dataset_loader(k, v) for k, v in DATASET_CONFIGS_GENERICS.items()},
-        "Lemhf14/EasyJailbreak_Datasets": partial(
-            load_multi_dataset,
-            name="Lemhf14/EasyJailbreak_Datasets",
-            sub_datasets=[
-                "AdvBench",
-                "ForbiddenQuestion",
-                "MJP",
-                "MaliciousInstruct",
-                "QuestionList",
-            ],
-        ),
-        "JailbreakV-28K/JailBreakV-28k": load_jailbreak_v28k,
-        "Local CSV": load_local_csv,
-        "Custom CSV": load_local_csv,
-    }
-
-    # Dynamic dataset loaders
-    dynamic_loaders = {
-        "AgenticBackend": lambda opts: dataset_from_iterator(
-            "AgenticBackend",
-            fine_tuned.Module(
-                opts["datasets"], tools_inbox=tools_inbox, opts=opts
-            ).apply(),
-            lazy=True,
-        ),
-        "Steganography": lambda opts: list(
-            StenographyTransformer(opts["datasets"]).transform()
-        ),
-        "llm-adaptive-attacks": lambda opts: dataset_from_iterator(
-            "llm-adaptive-attacks",
-            adaptive_attacks.Module(
-                opts["datasets"], tools_inbox=tools_inbox, opts=opts
-            ).apply(),
-        ),
-        "Garak": lambda opts: dataset_from_iterator(
-            "Garak",
-            garak_tool.Module(
-                opts["datasets"], tools_inbox=tools_inbox, opts=opts
-            ).apply(),
-            lazy=True,
-        ),
-        "Reinforcement Learning Optimization": lambda opts: dataset_from_iterator(
-            "Reinforcement Learning Optimization",
-            rl_model.Module(
-                opts["datasets"], tools_inbox=tools_inbox, opts=opts
-            ).apply(),
-            lazy=True,
-        ),
-        "InspectAI": lambda opts: dataset_from_iterator(
-            "InspectAI",
-            inspect_ai_tool.Module([], tools_inbox=tools_inbox).apply(),
-            lazy=True,
-        ),
-        "GPT fuzzer": lambda opts: [],
-    }
-
-    datasets = []
-    options = options or [dict(datasets=datasets) for _ in dataset_names]
-
-    # Load base datasets
-    for name, opts in zip(dataset_names, options):
-        if name not in dataset_loaders:
-            continue
+    for file in csv_files:
         try:
-            datasets.append(dataset_loaders[name]())
+            df = pd.read_csv(file)
         except Exception as e:
-            logger.error(f"Error loading {name}: {e}")
-
-    # Load dynamic datasets and apply transformations
-    for name, opts in zip(dataset_names, options):
-        if name not in dynamic_loaders:
+            logger.error(f"Error reading {file}: {e}")
             continue
-        logger.info(f"Loading dynamic dataset {name} {opts}")
-        opts["datasets"] = datasets
-        try:
-            dynamic_result = dynamic_loaders[name](opts)
-            datasets.extend(dynamic_result)
-        except Exception as e:
-            logger.exception(f"Error loading dynamic {name}: {e}")
+        # Check if 'prompt' column exists
+        if "prompt" in df.columns:
+            prompt_list.extend(df["prompt"].tolist())
+        else:
+            logger.warning(f"File {file} does not contain a 'prompt' column")
+    return ProbeDataset(
+        dataset_name="Local CSV",
+        metadata={"src": str(csv_files)},
+        prompts=prompt_list,
+        tokens=count_words_in_list(prompt_list),
+        approx_cost=0.0,
+    )
 
-    # Load csv datasets and apply transformations
-    for name, opts in zip(dataset_names, options):
-        if not name.endswith(".csv"):
-            continue
-        logger.info(f"Loading csv dataset {name} {opts}")
-        datasets.append(load_csv(name))
 
-    return datasets
+def dataset_from_iterator(name: str, iterator, lazy=False) -> list:
+    """Convert an iterator into a list of prompts and create a ProbeDataset
+    object.
+
+    Args:
+        name (str): The name of the dataset.
+        iterator (iterator): An iterator that generates prompts.
+
+    Returns:
+        list: A list containing a single ProbeDataset object.
+    """
+    prompts = list(iterator) if not lazy else iterator
+    tokens = count_words_in_list(prompts) if not lazy else 0
+    dataset = ProbeDataset(
+        dataset_name=name,
+        metadata={},
+        prompts=prompts,
+        tokens=tokens,
+        approx_cost=0.0,
+        lazy=lazy,
+    )
+    return [dataset]
+
+
+# TODO: refactor this abstraction
+
+dataset_map_generics = {
+    "simonycl/aya-23-8B_advbench_jailbreak": lambda: load_dataset_general(
+        dataset_name="simonycl/aya-23-8B_advbench_jailbreak",
+        column_mappings={"prompt": "prompt"},
+    ),
+    "acmc/jailbreaks_dataset_with_perplexity_bigcode_starcoder2-3b_bigcode_starcoder2-7b": lambda: load_dataset_general(
+        dataset_name="acmc/jailbreaks_dataset_with_perplexity_bigcode_starcoder2-3b_bigcode_starcoder2-7b"
+    ),
+    "karanxa/dolphin-jailbreak-finetuning-dataset": lambda: load_dataset_general(
+        dataset_name="karanxa/dolphin-jailbreak-finetuning-dataset",
+        column_mappings={"prompt": "text"},
+    ),
+    "karanxa/llama-2-jailbreak-dataset": lambda: load_dataset_general(
+        dataset_name="karanxa/llama-2-jailbreak-dataset",
+        column_mappings={"prompt": "text"},
+    ),
+    "karanxa/llama2-uncensored-jailbreak-dataset-finetuning": lambda: load_dataset_general(
+        dataset_name="karanxa/llama2-uncensored-jailbreak-dataset-finetuning",
+        column_mappings={"prompt": "text"},
+    ),
+    "liuyanchen1015/Llama-3.2-1B_jailbreak_responses": lambda: load_dataset_general(
+        dataset_name="liuyanchen1015/Llama-3.2-1B_jailbreak_responses",
+        column_mappings={"prompt": "jailbreak_prompt_text"},
+    ),
+    "liuyanchen1015/Llama-3.2-1B-Instruct_jailbreak_responses": lambda: load_dataset_general(
+        dataset_name="liuyanchen1015/Llama-3.2-1B-Instruct_jailbreak_responses",
+        column_mappings={"prompt": "jailbreak_prompt_text"},
+    ),
+    "liuyanchen1015/Llama-3.2-1B-Instruct_jailbreak_responses_with_judgment": lambda: load_dataset_general(
+        dataset_name="liuyanchen1015/Llama-3.2-1B-Instruct_jailbreak_responses_with_judgment",
+        column_mappings={"prompt": "jailbreak_prompt_text"},
+    ),
+    "jackhhao/jailbreak-classification": lambda: load_dataset_general(
+        dataset_name="jackhhao/jailbreak-classification",
+        column_mappings={"prompt": "prompt"},
+    ),
+    "markush1/LLM-Jailbreak-Classifier": lambda: load_dataset_general(
+        dataset_name="markush1/LLM-Jailbreak-Classifier",
+        column_mappings={"prompt": "prompt"},
+    ),
+    "walledai/JailbreakBench": lambda: load_dataset_general(
+        dataset_name="walledai/JailbreakBench", column_mappings={"prompt": "prompt"}
+    ),
+    "walledai/JailbreakHub": lambda: load_dataset_general(
+        dataset_name="walledai/JailbreakHub", column_mappings={"prompt": "prompt"}
+    ),
+    "Granther/evil-jailbreak": lambda: load_dataset_general(
+        dataset_name="Granther/evil-jailbreak", column_mappings={"prompt": "text"}
+    ),
+    "sevdeawesome/jailbreak_success": lambda: load_dataset_general(
+        dataset_name="sevdeawesome/jailbreak_success",
+        column_mappings={"prompt": "jailbreak_prompt_text"},
+    ),
+    "IDA-SERICS/Disaster-tweet-jailbreaking": lambda: load_dataset_general(
+        dataset_name="IDA-SERICS/Disaster-tweet-jailbreaking",
+        column_mappings={"prompt": "prompt_attack"},
+    ),
+    "GeorgeDaDude/Jailbreak_Complete_DS_labeled": lambda: load_dataset_general(
+        dataset_name="GeorgeDaDude/Jailbreak_Complete_DS_labeled",
+        column_mappings={"prompt": "question"},
+    ),
+    "dayone3nder/jailbreak_prompt_JBB_sft_trainset": lambda: load_dataset_general(
+        dataset_name="dayone3nder/jailbreak_prompt_JBB_sft_trainset",
+        column_mappings={"prompt": "prompt"},
+    ),
+    "dayone3nder/general_safe_mix_jailbreak_prompt_JBB_trainset": lambda: load_dataset_general(
+        dataset_name="dayone3nder/general_safe_mix_jailbreak_prompt_JBB_trainset",
+        column_mappings={"prompt": "prompt"},
+    ),
+}

agentic_security/probe_data/modules/adaptive_attacks.py

@@ -3,8 +3,7 @@ import io
 
 import httpx
 import pandas as pd
-
-from agentic_security.logutils import logger
+from loguru import logger
 
 url = "https://raw.githubusercontent.com/tml-epfl/llm-adaptive-attacks/main/harmful_behaviors/harmful_behaviors_pair.csv"
 

agentic_security/probe_data/modules/fine_tuned.py

@@ -3,8 +3,7 @@ import os
 import uuid as U
 
 import httpx
-
-from agentic_security.logutils import logger
+from loguru import logger
 
 AUTH_TOKEN: str = os.getenv("AS_TOKEN", "gh0-5f4a8ed2-37c6-4bd7-a0cf-7070eae8115b")
 
@@ -66,7 +65,7 @@ class Module:
                 return {}
 
     async def fetch_prompts(self) -> list[str]:
-        api_url = "https://mcp.metaheuristic.co/infer"
+        api_url = "https://edge.metaheuristic.co/infer"
         headers = {
             "Authorization": f"Bearer {AUTH_TOKEN}",
             "Content-Type": "application/json",

agentic_security/probe_data/modules/garak_tool.py

@@ -4,7 +4,7 @@ import json
 import os
 import subprocess
 
-from agentic_security.logutils import logger
+from loguru import logger
 
 # TODO: add probes modules
 

agentic_security/probe_data/modules/inspect_ai_tool.py

@@ -2,7 +2,7 @@ import asyncio
 import importlib.util
 import os
 
-from agentic_security.logutils import logger
+from loguru import logger
 
 inspect_ai_task = (
     __file__.replace("inspect_ai_tool.py", "inspect_ai_task.py")

agentic_security/probe_data/modules/rl_model.py

@@ -8,8 +8,7 @@ from typing import Deque
 
 import numpy as np
 import requests
-
-from agentic_security.logutils import logger
+from loguru import logger
 
 AUTH_TOKEN: str = os.getenv("AS_TOKEN", "gh0-5f4a8ed2-37c6-4bd7-a0cf-7070eae8115b")
 
@@ -216,7 +215,7 @@ class Module:
         self.run_id = U.uuid4().hex
         self.batch_size = self.opts.get("batch_size", 500)
         self.rl_model = CloudRLPromptSelector(
-            prompt_groups, "https://mcp.metaheuristic.co", run_id=self.run_id
+            prompt_groups, "https://edge.metaheuristic.co", run_id=self.run_id
         )
 
     async def apply(self):

agentic_security/probe_data/modules/test_rl_model.py

@@ -36,7 +36,7 @@ def mock_requests() -> Mock:
 def mock_rl_selector() -> Mock:
     return CloudRLPromptSelector(
         dataset_prompts,
-        api_url="https://mcp.metaheuristic.co",
+        api_url="https://edge.metaheuristic.co",
     )
 
 
@@ -94,7 +94,7 @@ class TestCloudRLPromptSelector:
     def test_select_next_prompt_success_service(self, dataset_prompts):
         selector = CloudRLPromptSelector(
             dataset_prompts,
-            api_url="https://mcp.metaheuristic.co",
+            api_url="https://edge.metaheuristic.co",
         )
         next_prompt = selector.select_next_prompt(
             "How does RL work?", passed_guard=True

agentic_security/probe_data/msj_data.py

@@ -1,6 +1,6 @@
 from dataclasses import dataclass
 
-from cache_to_disk import cache_to_disk  # noqa
+from cache_to_disk import cache_to_disk
 
 
 # TODO: refactor this class to use from .data
@@ -22,7 +22,7 @@ class ProbeDataset:
         }
 
 
-# @cache_to_disk(n_days_to_cache=1)
+@cache_to_disk()
 def load_dataset_generic(name, getter=lambda x: x["train"]["prompt"]):
     from datasets import load_dataset
 

agentic_security/refusal_classifier/test_model.py

@@ -3,7 +3,7 @@ from unittest.mock import patch
 import pandas as pd
 import pytest
 
-from agentic_security.refusal_classifier.model import RefusalClassifier
+from .model import RefusalClassifier
 
 
 @pytest.fixture

agentic_security/report_chart.py

@@ -7,61 +7,9 @@ import pandas as pd
 from matplotlib.cm import ScalarMappable
 from matplotlib.colors import LinearSegmentedColormap, Normalize
 
-from agentic_security.logutils import logger
 
-from .primitives import Table
-
-
-def plot_security_report(table: Table) -> io.BytesIO:
-    """
-    Generates a polar plot representing the security report based on the given data.
-
-    Args:
-        table (Table): The input data table containing security metrics.
-
-    Returns:
-        io.BytesIO: A buffer containing the generated plot image in PNG format.
-                    Returns an empty buffer in case of an error.
-    """
-    try:
-        return _plot_security_report(table=table)
-    except (TypeError, ValueError, OverflowError, IndexError, Exception) as e:
-        logger.error(f"Error in generating the security report: {e} {table}")
-    return io.BytesIO()
-
-
-def generate_identifiers(data: pd.DataFrame) -> list[str]:
-    """
-    Generates unique identifiers for the given dataset.
-
-    Args:
-        data (pd.DataFrame): A pandas DataFrame containing security-related data.
-
-    Returns:
-        list[str]: A list of generated identifiers. Returns a list with an empty string in case of an error.
-    """
-    return _generate_identifiers(data=data)
-
-
-def _plot_security_report(table: Table) -> io.BytesIO:
-    """
-    Generates a polar plot-based security report visualizing the failure rates for different modules.
-
-    This function processes the input data, sorts it by failure rate, and generates a polar plot
-    where each bar represents the failure rate for a specific module. The plot includes identifiers,
-    color-coding based on token count, failure rate values on the bars, and a table listing the modules
-    and their corresponding failure rates.
-
-    Args:
-        table (Table): A table-like structure (e.g., pandas DataFrame) containing security report data
-                        with columns for failure rate, tokens, and modules.
-
-    Returns:
-        io.BytesIO: A buffer containing the generated plot image in PNG format.
-    """
+def plot_security_report(table):
     # Data preprocessing
-    logger.info("Data preprocessing started.")
-
     data = pd.DataFrame(table)
 
     # Sort by failure rate and reset index
@@ -72,10 +20,10 @@ def _plot_security_report(table: Table) -> io.BytesIO:
     fig, ax = plt.subplots(figsize=(12, 10), subplot_kw={"projection": "polar"})
     fig.set_facecolor("#f0f0f0")
     ax.set_facecolor("#f0f0f0")
-    logger.info("Plot setup complete.")
 
     # Styling parameters
     colors = ["#6C5B7B", "#C06C84", "#F67280", "#F8B195"][::-1]  # Pastel palette
+    # colors = ["#440154", "#3b528b", "#21908c", "#5dc863"]  # Viridis-inspired palette
     cmap = LinearSegmentedColormap.from_list("custom", colors, N=256)
     norm = Normalize(vmin=data["tokens"].min(), vmax=data["tokens"].max())
 
@@ -126,10 +74,7 @@ def _plot_security_report(table: Table) -> io.BytesIO:
 
     # Title and caption
     fig.suptitle(
-        "Security Report for Different Modules",
-        fontsize=16,
-        fontweight="bold",
-        y=1.02,
+        "Security Report for Different Modules", fontsize=16, fontweight="bold", y=1.02
     )
     caption = "Report generated by https://github.com/msoedov/agentic_security"
     fig.text(
@@ -167,12 +112,17 @@ def _plot_security_report(table: Table) -> io.BytesIO:
             data["identifier"], data["failureRate"], data["module"]
         )
     ]
-    table = ax.table(cellText=table_data, loc="right", cellLoc="left")
+    table = ax.table(
+        cellText=table_data,
+        loc="right",
+        cellLoc="left",
+    )
     table.auto_set_font_size(False)
     table.set_fontsize(8)
 
     # Adjust table style
     table.scale(1, 0.7)
+
     for (row, col), cell in table.get_celld().items():
         cell.set_edgecolor("none")
         cell.set_facecolor("#f0f0f0" if row % 2 == 0 else "#e0e0e0")
@@ -182,33 +132,17 @@ def _plot_security_report(table: Table) -> io.BytesIO:
             cell.set_text_props(fontweight="bold")
 
     # Adjust layout and save
+
     plt.tight_layout()
     buf = io.BytesIO()
     plt.savefig(buf, format="png", dpi=300, bbox_inches="tight")
     plt.close(fig)
     buf.seek(0)
-    logger.info("Report successfully generated and saved to buffer.")
     return buf
 
 
-def _generate_identifiers(data: pd.DataFrame) -> list[str]:
-    """
-    Generates a list of unique identifiers for each row in the given DataFrame.
-
-    The identifiers are based on the English alphabet, with each identifier consisting
-    of a letter followed by a number. The letter represents the "group" of identifiers
-    (using a letter from A to Z) and the number is a counter within that group. For example:
-    - A1, A2, ..., A26, B1, B2, ..., Z1, Z2, ...
-
-    Args:
-        data (pd.DataFrame): The input DataFrame containing data for which identifiers
-                             are to be generated.
-
-    Returns:
-        list[str]: A list of unique identifiers as strings, one for each row in the DataFrame.
-    """
+def generate_identifiers(data):
     data_length = len(data)
-
     alphabet = string.ascii_uppercase
     num_letters = len(alphabet)
 

agentic_security/routes/_specs.py

@@ -1,397 +0,0 @@
-_SPECS = [
-    """POST ${SELF_URL}/v1/self-probe
-Authorization: Bearer XXXXX
-Content-Type: application/json
-
-{
-"prompt": "<<PROMPT>>"
-}
-
-""",
-    """POST https://api.openai.com/v1/chat/completions
-Authorization: Bearer $OPENAI_API_KEY
-Content-Type: application/json
-
-{
-"model": "gpt-3.5-turbo",
-"messages": [{"role": "user", "content": "<<PROMPT>>"}],
-"temperature": 0.7
-}
-""",
-    """
-POST https://api.deepseek.com/chat/completions
-Authorization: Bearer $DEEPSEEK_API_KEY
-Content-Type: application/json
-
-{
-  "model": "deepseek-chat",
-  "messages": [
-    {"role": "system", "content": "You are a helpful assistant."},
-    {"role": "user", "content": "<<PROMPT>>"}
-  ],
-  "stream": false
-}
-""",
-    """POST https://api.replicate.com/v1/models/mistralai/mixtral-8x7b-instruct-v0.1/predictions
-Authorization: Bearer $APIKEY
-Content-Type: application/json
-
-{
-"input": {
-"top_k": 50,
-"top_p": 0.9,
-"prompt": "Write a bedtime story about neural networks I can read to my toddler",
-"temperature": 0.6,
-"max_new_tokens": 1024,
-"prompt_template": "<s>[INST] <<PROMPT>> [/INST] ",
-"presence_penalty": 0,
-"frequency_penalty": 0
-}
-}
-""",
-    """POST https://api.groq.com/v1/request_manager/text_completion
-Authorization: Bearer $APIKEY
-Content-Type: application/json
-
-{
-"model_id": "codellama-34b",
-"system_prompt": "You are helpful and concise coding assistant",
-"user_prompt": "<<PROMPT>>"
-}
-""",
-    """POST https://api.together.xyz/v1/chat/completions
-Authorization: Bearer $TOGETHER_API_KEY
-Content-Type: application/json
-
-{
-"model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-"messages": [
-{"role": "system", "content": "You are an expert travel guide"},
-{"role": "user", "content": "<<PROMPT>>"}
-]
-}
-""",
-    """POST ${SELF_URL}/v1/self-probe-image
-Authorization: Bearer XXXXX
-Content-Type: application/json
-
-[
-    {
-      "role": "user",
-      "content": [
-        {
-          "type": "text",
-          "text": "What is in this image?",
-        },
-        {
-          "type": "image_url",
-          "image_url": {
-            "url":  f"data:image/jpeg;base64,{<<BASE64_IMAGE>>}"
-          },
-        },
-      ],
-    }
-]
-""",
-    """POST ${SELF_URL}/v1/self-probe-file
-Authorization: Bearer $API_KEY
-Content-Type: multipart/form-data
-
-{
-  "file": "@./sample_audio.m4a",
-  "model": "whisper-large-v3"
-}
-""",
-    """POST https://api.gemini.com/v1/generate
-Authorization: Bearer $GEMINI_API_KEY
-Content-Type: application/json
-
-{
-  "model": "gemini-latest",
-  "prompt": "<<PROMPT>>",
-  "temperature": 0.8,
-  "max_tokens": 150,
-  "top_p": 1.0,
-  "frequency_penalty": 0,
-  "presence_penalty": 0
-}
-""",
-    """POST https://api.anthropic.com/v1/complete
-Authorization: Bearer $ANTHROPIC_API_KEY
-Content-Type: application/json
-
-{
-  "model": "claude-v1.3",
-  "prompt": "<<PROMPT>>",
-  "temperature": 0.7,
-  "max_tokens_to_sample": 256,
-  "stop_sequences": ["\n\nHuman:"]
-}
-""",
-    """POST https://api.cohere.ai/generate
-Authorization: Bearer $COHERE_API_KEY
-Content-Type: application/json
-
-{
-  "model": "command-xlarge-nightly",
-  "prompt": "<<PROMPT>>",
-  "max_tokens": 300,
-  "temperature": 0.75,
-  "k": 0,
-  "p": 0.75
-}
-""",
-    """POST https://<<RESOURCE_NAME>>.openai.azure.com/openai/deployments/<<DEPLOYMENT_NAME>>/completions?api-version=2023-06-01-preview
-Authorization: Bearer $AZURE_API_KEY
-Content-Type: application/json
-
-{
-  "prompt": "<<PROMPT>>",
-  "max_tokens": 150,
-  "temperature": 0.7,
-  "top_p": 0.9,
-  "frequency_penalty": 0,
-  "presence_penalty": 0
-}
-""",
-    """POST https://api.assemblyai.com/v2/transcript
-Authorization: Bearer $ASSEMBLY_API_KEY
-Content-Type: application/json
-
-{
-  "audio_url": "<<AUDIO_FILE_URL>>"
-}
-""",
-    """POST https://api.openrouter.ai/v1/chat/completions
-Authorization: Bearer $OPENROUTER_API_KEY
-Content-Type: application/json
-
-{
-  "model": "openrouter-latest",
-  "prompt": "<<PROMPT>>",
-  "temperature": 0.7,
-  "max_tokens": 150,
-  "top_p": 0.9,
-  "frequency_penalty": 0,
-  "presence_penalty": 0
-}
-""",
-]
-
-
-LLM_SPECS = [
-    """POST ${SELF_URL}/v1/self-probe
-Authorization: Bearer XXXXX
-Content-Type: application/json
-
-{
-"prompt": "<<PROMPT>>"
-}
-
-""",
-    """POST https://api.openai.com/v1/chat/completions
-Authorization: Bearer $OPENAI_API_KEY
-Content-Type: application/json
-
-{
-"model": "gpt-3.5-turbo",
-"messages": [{"role": "user", "content": "<<PROMPT>>"}],
-"temperature": 0.7
-}
-""",
-    """
-POST https://api.deepseek.com/chat/completions
-Authorization: Bearer $DEEPSEEK_API_KEY
-Content-Type: application/json
-
-{
-  "model": "deepseek-chat",
-  "messages": [
-    {"role": "system", "content": "You are a helpful assistant."},
-    {"role": "user", "content": "<<PROMPT>>"}
-  ],
-  "stream": false
-}
-""",
-    """POST https://api.replicate.com/v1/models/mistralai/mixtral-8x7b-instruct-v0.1/predictions
-Authorization: Bearer $APIKEY
-Content-Type: application/json
-
-{
-"input": {
-"top_k": 50,
-"top_p": 0.9,
-"prompt": "Write a bedtime story about neural networks I can read to my toddler",
-"temperature": 0.6,
-"max_new_tokens": 1024,
-"prompt_template": "<s>[INST] <<PROMPT>> [/INST] ",
-"presence_penalty": 0,
-"frequency_penalty": 0
-}
-}
-""",
-    """POST https://api.groq.com/v1/request_manager/text_completion
-Authorization: Bearer $APIKEY
-Content-Type: application/json
-
-{
-"model_id": "codellama-34b",
-"system_prompt": "You are helpful and concise coding assistant",
-"user_prompt": "<<PROMPT>>"
-}
-""",
-    """POST https://api.together.xyz/v1/chat/completions
-Authorization: Bearer $TOGETHER_API_KEY
-Content-Type: application/json
-
-{
-"model": "mistralai/Mixtral-8x7B-Instruct-v0.1",
-"messages": [
-{"role": "system", "content": "You are an expert travel guide"},
-{"role": "user", "content": "<<PROMPT>>"}
-]
-}
-""",
-    """POST ${SELF_URL}/v1/self-probe-image
-Authorization: Bearer XXXXX
-Content-Type: application/json
-
-[
-    {
-      "role": "user",
-      "content": [
-        {
-          "type": "text",
-          "text": "What is in this image?",
-        },
-        {
-          "type": "image_url",
-          "image_url": {
-            "url":  f"data:image/jpeg;base64,{<<BASE64_IMAGE>>}"
-          },
-        },
-      ],
-    }
-]
-""",
-    """POST ${SELF_URL}/v1/self-probe-file
-Authorization: Bearer $API_KEY
-Content-Type: multipart/form-data
-
-{
-  "file": "@./sample_audio.m4a",
-  "model": "whisper-large-v3"
-}
-""",
-    """POST https://api.gemini.com/v1/generate
-Authorization: Bearer $GEMINI_API_KEY
-Content-Type: application/json
-
-{
-  "model": "gemini-latest",
-  "prompt": "<<PROMPT>>",
-  "temperature": 0.8,
-  "max_tokens": 150,
-  "top_p": 1.0,
-  "frequency_penalty": 0,
-  "presence_penalty": 0
-}
-""",
-    """POST https://api.anthropic.com/v1/complete
-Authorization: Bearer $ANTHROPIC_API_KEY
-Content-Type: application/json
-
-{
-  "model": "claude-v1.3",
-  "prompt": "<<PROMPT>>",
-  "temperature": 0.7,
-  "max_tokens_to_sample": 256,
-  "stop_sequences": ["\n\nHuman:"]
-}
-""",
-    """POST https://api.cohere.ai/generate
-Authorization: Bearer $COHERE_API_KEY
-Content-Type: application/json
-
-{
-  "model": "command-xlarge-nightly",
-  "prompt": "<<PROMPT>>",
-  "max_tokens": 300,
-  "temperature": 0.75,
-  "k": 0,
-  "p": 0.75
-}
-""",
-    """POST https://<<RESOURCE_NAME>>.openai.azure.com/openai/deployments/<<DEPLOYMENT_NAME>>/completions?api-version=2023-06-01-preview
-Authorization: Bearer $AZURE_API_KEY
-Content-Type: application/json
-
-{
-  "prompt": "<<PROMPT>>",
-  "max_tokens": 150,
-  "temperature": 0.7,
-  "top_p": 0.9,
-  "frequency_penalty": 0,
-  "presence_penalty": 0
-}
-""",
-    """POST https://api.assemblyai.com/v2/transcript
-Authorization: Bearer $ASSEMBLY_API_KEY
-Content-Type: application/json
-
-{
-  "audio_url": "<<AUDIO_FILE_URL>>"
-}
-""",
-    """POST https://api.openrouter.ai/v1/chat/completions
-Authorization: Bearer $OPENROUTER_API_KEY
-Content-Type: application/json
-
-{
-  "model": "openrouter-latest",
-  "prompt": "<<PROMPT>>",
-  "temperature": 0.7,
-  "max_tokens": 150,
-  "top_p": 0.9,
-  "frequency_penalty": 0,
-  "presence_penalty": 0
-}
-""",
-]
-
-
-LLM_CONFIGS = [
-    {
-        "name": "Custom API",
-        "prompts": 40000,
-        "customInstructions": "Requires api spec",
-        "logo": "/icons/myshell.png",
-    },
-    {"name": "Open AI", "prompts": 24000, "logo": "/icons/openai.png"},
-    {"name": "Deepseek v1", "prompts": 24000, "logo": "/icons/deepseek.png"},
-    {"name": "Replicate", "prompts": 40000, "logo": "/icons/replicate.png"},
-    {"name": "Groq", "prompts": 40000, "logo": "/icons/groq.png"},
-    {"name": "Together.ai", "prompts": 40000, "logo": "/icons/together.png"},
-    {
-        "name": "Custom API Image",
-        "prompts": 40000,
-        "customInstructions": "Requires api spec",
-        "modality": "Image",
-        "logo": "/icons/myshell.png",
-    },
-    {
-        "name": "Custom API Files",
-        "prompts": 40000,
-        "customInstructions": "Requires api spec",
-        "modality": "Files",
-        "logo": "/icons/myshell.png",
-    },
-    {"name": "Gemini", "prompts": 40000, "logo": "/icons/gemini.png"},
-    {"name": "Claude", "prompts": 40000, "logo": "/icons/claude.png"},
-    {"name": "Cohere", "prompts": 40000, "logo": "/icons/cohere.png"},
-    {"name": "Azure OpenAI", "prompts": 40000, "logo": "/icons/azureai.png"},
-    {"name": "assemblyai", "prompts": 40000, "logo": "/icons/myshell.png"},
-    {"name": "OpenRouter.ai", "prompts": 40000, "logo": "/icons/openrouter.png"},
-]
-
-LLM_SPECS = [dict(spec=spec, **d) for spec, d in zip(_SPECS, LLM_CONFIGS)]

agentic_security/routes/probe.py

@@ -6,7 +6,6 @@ from fastapi.responses import JSONResponse
 from ..primitives import FileProbeResponse, Probe
 from ..probe_actor.refusal import REFUSAL_MARKS
 from ..probe_data import REGISTRY
-from ._specs import LLM_SPECS
 
 router = APIRouter()
 
@@ -74,12 +73,6 @@ async def data_config():
     return [m for m in REGISTRY]
 
 
-@router.get("/v1/llm-specs", response_model=list)
-def get_llm_specs():
-    """Returns the LLM API specifications."""
-    return LLM_SPECS
-
-
 @router.get("/health")
 async def health_check():
     """Health check endpoint."""

agentic_security/routes/proxy.py

@@ -2,8 +2,7 @@ import random
 from asyncio import Event
 
 from fastapi import APIRouter
-
-from agentic_security.logutils import logger
+from loguru import logger
 
 from ..core.app import get_current_run, get_tools_inbox
 from ..primitives import CompletionRequest, Settings

agentic_security/routes/scan.py

@@ -1,6 +1,4 @@
-from collections.abc import Generator
 from datetime import datetime
-from typing import Any
 
 from fastapi import (
     APIRouter,
@@ -13,11 +11,9 @@ from fastapi import (
 )
 from fastapi.responses import StreamingResponse
 
-from agentic_security.logutils import logger
-
 from ..core.app import get_stop_event, get_tools_inbox, set_current_run
 from ..dependencies import InMemorySecrets, get_in_memory_secrets
-from ..http_spec import InvalidHTTPSpecError, LLMSpec
+from ..http_spec import LLMSpec
 from ..primitives import LLMInfo, Scan
 from ..probe_actor import fuzzer
 
@@ -27,16 +23,9 @@ router = APIRouter()
 @router.post("/verify")
 async def verify(
     info: LLMInfo, secrets: InMemorySecrets = Depends(get_in_memory_secrets)
-) -> dict[str, int | str | float]:
+):
     spec = LLMSpec.from_string(info.spec)
-    try:
-        r = await spec.verify()
-    except InvalidHTTPSpecError as e:
-        raise HTTPException(status_code=400, detail=str(e))
-    except Exception as e:
-        logger.exception(e)
-        raise HTTPException(status_code=400, detail=str(e))
-
+    r = await spec.verify()
     if r.status_code >= 400:
         raise HTTPException(status_code=r.status_code, detail=r.text)
     return dict(
@@ -47,7 +36,7 @@ async def verify(
     )
 
 
-def streaming_response_generator(scan_parameters: Scan) -> Generator[str, Any, None]:
+def streaming_response_generator(scan_parameters: Scan):
     request_factory = LLMSpec.from_string(scan_parameters.llmSpec)
     set_current_run(request_factory)
 
@@ -68,7 +57,7 @@ async def scan(
     scan_parameters: Scan,
     background_tasks: BackgroundTasks,
     secrets: InMemorySecrets = Depends(get_in_memory_secrets),
-) -> StreamingResponse:
+):
     scan_parameters.with_secrets(secrets)
     return StreamingResponse(
         streaming_response_generator(scan_parameters), media_type="application/json"
@@ -76,7 +65,7 @@ async def scan(
 
 
 @router.post("/stop")
-async def stop_scan() -> dict[str, str]:
+async def stop_scan():
     get_stop_event().set()
     return {"status": "Scan stopped"}
 
@@ -90,7 +79,7 @@ async def scan_csv(
     maxBudget: int = Query(10_000),
     enableMultiStepAttack: bool = Query(False),
     secrets: InMemorySecrets = Depends(get_in_memory_secrets),
-) -> StreamingResponse:
+):
     # TODO: content dataset to fuzzer
     content = await file.read()  # noqa
     llm_spec = await llmSpec.read()

agentic_security/routes/telemetry.py

@@ -1,8 +1,7 @@
 import sentry_sdk
+from loguru import logger
 from sentry_sdk.integrations.logging import ignore_logger
 
-from agentic_security.logutils import logger
-
 from ..primitives import Settings
 
 

agentic_security/routes/test_health.py

@@ -1,6 +1,6 @@
 from fastapi.testclient import TestClient
 
-from agentic_security.app import app
+from ..app import app
 
 
 def test_health_check():

agentic_security/routes/test_probe.py

@@ -5,10 +5,10 @@ import httpx
 import pytest
 from fastapi.testclient import TestClient
 
-from agentic_security.app import app
-from agentic_security.primitives import Probe
-from agentic_security.probe_actor.refusal import REFUSAL_MARKS
-from agentic_security.probe_data import REGISTRY
+from ..app import app
+from ..primitives import Probe
+from ..probe_actor.refusal import REFUSAL_MARKS
+from ..probe_data import REGISTRY
 
 client = TestClient(app)
 

agentic_security/routes/test_report.py

@@ -4,7 +4,7 @@ from unittest.mock import patch
 import pytest
 from fastapi.testclient import TestClient
 
-from agentic_security.routes.report import router
+from .report import router
 
 client = TestClient(router)
 

agentic_security/routes/test_static.py

@@ -4,8 +4,8 @@ import pytest
 from fastapi import HTTPException
 from fastapi.testclient import TestClient
 
-from agentic_security.primitives import Settings
-from agentic_security.routes.static import get_static_file, router
+from ..primitives import Settings
+from .static import get_static_file, router
 
 client = TestClient(router)
 

agentic_security/static/base.js

@@ -101,7 +101,7 @@ Content-Type: application/json
 ]
 `,
   `POST ${SELF_URL}/v1/self-probe-file
-Authorization: Bearer $API_KEY
+Authorization: Bearer $GROQ_API_KEY
 Content-Type: multipart/form-data
 
 {
@@ -204,9 +204,8 @@ let LLM_CONFIGS = [
   { name: 'Claude', prompts: 40000, logo: '/icons/claude.png' },
   { name: 'Cohere', prompts: 40000, logo: '/icons/cohere.png' },
   { name: 'Azure OpenAI', prompts: 40000, logo: '/icons/azureai.png' },
-  { name: 'assemblyai', prompts: 40000, logo: fallbackIcon },
   { name: 'OpenRouter.ai', prompts: 40000, logo: '/icons/openrouter.png' },
-
+  { name: 'assemblyai', prompts: 40000, logo: fallbackIcon },
 ];
 function has_image(spec) {
   return spec.includes('<<BASE64_IMAGE>>');

agentic_security/static/index.html

@@ -63,32 +63,6 @@
           </div>
         </section>
       </main>
-      <div class="fixed top-6 right-6 z-50 space-y-3">
-        <transition-group name="toast">
-            <div
-                v-for="toast in toasts"
-                :key="toast.id"
-                class="flex items-center p-3 rounded-xl shadow-xl text-white max-w-md animate-toast-in border border-opacity-30"
-                :class="{
-                    'bg-success-toast border-accent-green': toast.type === 'success',
-                    'bg-error-toast border-accent-red': toast.type === 'error',
-                    'bg-info-toast border-accent-orange': toast.type === 'info'
-                }"
-            >
-                <span class="flex-1 font-medium tracking-wide text-sm">{{ toast.message }}</span>
-                <button
-                    @click="removeToast(toast.id)"
-                    class="ml-3 focus:outline-none hover:opacity-80 transition-opacity"
-                >
-                    <svg class="w-4 h-4" fill="none" stroke="currentColor" viewBox="0 0 24 24">
-                        <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M6 18L18 6M6 6l12 12" />
-                    </svg>
-                </button>
-            </div>
-        </transition-group>
-    </div>
-
-
       <main class="max-w-6xl mx-auto space-y-8">
         <section class="bg-dark-card rounded-lg p-6 shadow-lg" v-show="false">
           <h2 class="text-2xl font-bold mb-4">Select a Config</h2>

agentic_security/static/main.js

@@ -25,8 +25,6 @@ var app = new Vue({
         showModules: false,
         showLogs: false,
         showConsentModal: true,
-        toasts: [], // Array to store toast notifications
-        toastTimeout: 3000, // Duration in milliseconds (3 seconds)
         statusDotClass: 'bg-gray-500', // Default status dot class
         statusText: 'Verified', // Default status text
         statusClass: 'bg-green-500 text-dark-bg', // Default status class
@@ -95,36 +93,21 @@ var app = new Vue({
 
     },
     methods: {
-        showToast(message, type = 'success') {
-            const id = Date.now(); // Unique ID for each toast
-            this.toasts.push({ id, message, type });
-
-            // Automatically remove toast after timeout
-            setTimeout(() => {
-                this.removeToast(id);
-            }, this.toastTimeout);
-        },
-
-        removeToast(id) {
-            this.toasts = this.toasts.filter(toast => toast.id !== id);
-        },
         focusTextarea() {
             this.isFocused = true;
-            // Remove 'self' assignment if not used elsewhere
+            self = this.$refs;
             this.$nextTick(() => {
-                this.$refs.textarea.focus();
-                this.adjustHeight({ target: this.$refs.textarea });
+                // Focus the textarea after rendering
+                self.textarea.focus();
+                this.adjustHeight({ target: self.textarea });
             });
-            // Correct the event listener to use handleOutsideClick
-            document.addEventListener("mousedown", this.handleOutsideClick);
+            document.addEventListener("mousedown", this.handleClickOutside);
+
         },
         handleOutsideClick(event) {
-            if (!this.$refs.textarea) {
-                return
-            }
-            if (!this.$refs.textarea.contains(event.target)) {
+            if (!this.$refs.container.contains(event.target)) {
                 this.isFocused = false;
-                document.removeEventListener("mousedown", this.handleOutsideClick);
+                document.removeEventListener("mousedown", this.handleClickOutside);
             }
         },
         unfocusTextarea() {
@@ -132,12 +115,7 @@ var app = new Vue({
         },
         acceptConsent() {
             this.showConsentModal = false; // Close the modal
-
-            try {
-                localStorage.setItem('consentGiven', 'true'); // Save consent to local storage
-            } catch (e) {
-                this.showToast('Failed to save consent', 'error'); // Show error if saving fails
-            }
+            localStorage.setItem('consentGiven', 'true'); // Save consent to local storage
         },
 
         saveStateToLocalStorage() {
@@ -178,7 +156,6 @@ var app = new Vue({
             this.integrationVerified = false;
             this.showResetConfirmation = false;
             this.enableMultiStepAttack = false;
-            this.showToast('All settings have been reset to default', 'info');
         },
         confirmResetState() {
             this.showResetConfirmation = true;
@@ -217,39 +194,31 @@ var app = new Vue({
                 spec: this.modelSpec,
             };
             let startTime = performance.now(); // Capture start time
-
-            try {
-                const response = await fetch(`${SELF_URL}/verify`, {
-                    method: 'POST',
-                    headers: {
-                        'Content-Type': 'application/json',
-                    },
-                    body: JSON.stringify(payload),
-                });
-
-                let r = await response.json();
-
-                let endTime = performance.now(); // Capture end time
-                let latency = ((endTime - startTime) / 1000).toFixed(3); // Calculate latency in milliseconds
-                this.latency = latency;
-
-                if (!response.ok) {
-                    this.updateStatusDot(false);
-                    this.errorMsg = 'Integration verification failed:' + JSON.stringify(r);
-                    this.showToast('Integration verification failed', 'error');
-                } else {
-                    this.errorMsg = '';
-                    this.updateStatusDot(true);
-                    this.okMsg = 'Integration verified';
-                    this.showToast('Integration verified successfully', 'success');
-                    this.integrationVerified = true;
-                }
-            } catch (error) {
+            const response = await fetch(`${SELF_URL}/verify`, {
+                method: 'POST',
+                headers: {
+                    'Content-Type': 'application/json',
+                },
+                body: JSON.stringify(payload),
+            });
+            console.log(response);
+            let r = await response.json();
+            let endTime = performance.now(); // Capture end time
+            let latency = endTime - startTime; // Calculate latency in milliseconds
+            latency = latency.toFixed(3) / 1000; // Round to 2 decimal places
+            this.latency = latency;
+            if (!response.ok) {
+                this.updateStatusDot(false);
+                this.errorMsg = 'Integration verification failed:' + JSON.stringify(r);
+            } else {
+                this.errorMsg = '';
                 this.updateStatusDot(true);
-                this.errorMsg = 'Server unreachable';
-                this.showToast('Network error', 'error');
-            }
+                this.okMsg = 'Integration verified';
+                this.integrationVerified = true;
+                // console.log('Integration verified', this.integrationVerified);
+                // this.$forceUpdate();
 
+            }
             this.saveStateToLocalStorage();
         },
         loadConfigs: async function () {
@@ -271,7 +240,6 @@ var app = new Vue({
             this.errorMsg = '';
             this.okMsg = '';
             this.integrationVerified = false;
-            this.showToast(`Config ${index + 1} selected`, 'info');
         },
         toggleModules() {
             this.showModules = !this.showModules;
@@ -359,7 +327,6 @@ var app = new Vue({
                 return
             }
             console.log('New row');
-            this.showToast('New module', 'success');
             let payload = {
                 table: this.mainTable,
             };
@@ -470,8 +437,6 @@ var app = new Vue({
                     }
                 });
             }
-            this.scanRunning = false;
-            this.showToast('Scan finished successfully', 'success');
             this.saveStateToLocalStorage();
 
         }

agentic_security/static/partials/head.html

@@ -19,7 +19,7 @@
               technopollas: ['Technopollas', 'sans-serif'],
             },
             colors: {
-                dark: {
+                t1: {
                 bg: '#0D0D0D', // Jet Black
                 card: '#1A1A1A', // Dark Carbon Fiber
                 text: '#FFFFFF',
@@ -30,7 +30,7 @@
                   yellow: '#2E4053', // Jet Black
                 },
               },
-              dark1: {
+              dark: {
                 bg: '#121212',
                 card: '#1E1E1E',
                 text: '#FFFFFF',
@@ -86,7 +86,6 @@
         }
       }
     </script>
-
     <style>
 .scrollbar-hide::-webkit-scrollbar {
   display: none;
@@ -97,55 +96,4 @@
 }
 
     </style>
-<style>
-/* Toast-specific colors */
-.bg-success-toast {
-    background: #1C3F74
-}
-
-.bg-error-toast {
-    background: #85144B
-}
-
-.bg-info-toast {
-    background: #FFC300
-}
-
-.border-accent-green, .border-accent-red, .border-accent-orange {
-    border-color: rgba(255, 255, 255, 0.1); /* Subtle white border for depth */
-}
-
-/* Animation */
-.animate-toast-in {
-    animation: tSlideIn 0.4s cubic-bezier(0.25, 0.1, 0.25, 1);
-}
-
-@keyframes tSlideIn {
-    from {
-        transform: translateX(120%) scale(0.95);
-        opacity: 0;
-    }
-    to {
-        transform: translateX(0) scale(1);
-        opacity: 1;
-    }
-}
-
-.toast-enter-active,
-.toast-leave-active {
-    transition: all 0.4s cubic-bezier(0.25, 0.1, 0.25, 1);
-}
-
-.toast-enter-from,
-.toast-leave-to {
-    transform: translateX(120%) scale(0.95);
-    opacity: 0;
-}
-
-/* Hover effect */
-[toast-type]:hover {
-    transform: translateY(-2px);
-    box-shadow: 0 6px 20px rgba(0, 0, 0, 0.6);
-}
-</style>
   </head>

agentic_security/test_lib.py

@@ -8,7 +8,7 @@ import time
 import pytest
 
 import agentic_security.test_spec_assets as test_spec_assets
-from agentic_security.lib import SecurityScanner
+from agentic_security.lib import AgenticSecurity
 
 
 def has_module(module_name):
@@ -57,7 +57,7 @@ class TestLibraryLevel:
         maxBudget = 1000000
         max_th = 0.3
         datasets = make_test_registry()
-        result = SecurityScanner.scan(llmSpec, maxBudget, datasets, max_th)
+        result = AgenticSecurity.scan(llmSpec, maxBudget, datasets, max_th)
         assert isinstance(result, dict)
         print(result)
         assert len(result) in [0, 1]
@@ -68,7 +68,7 @@ class TestLibraryLevel:
         maxBudget = 1000
         max_th = 0.3
         datasets = make_test_registry()
-        result = SecurityScanner.scan(
+        result = AgenticSecurity.scan(
             llmSpec, maxBudget, datasets, max_th, enableMultiStepAttack=True
         )
         assert isinstance(result, dict)
@@ -93,7 +93,7 @@ class TestLibraryLevel:
                 "opts": {"port": 9094},
             },
         ]
-        result = SecurityScanner.scan(llmSpec, maxBudget, datasets, max_th)
+        result = AgenticSecurity.scan(llmSpec, maxBudget, datasets, max_th)
         assert isinstance(result, dict)
         print(result)
         assert len(result) in [0, 1]
@@ -120,7 +120,7 @@ class TestLibraryLevel:
                 "modality": "text",
             },
         ]
-        result = SecurityScanner.scan(llmSpec, maxBudget, datasets, max_th)
+        result = AgenticSecurity.scan(llmSpec, maxBudget, datasets, max_th)
         assert isinstance(result, dict)
         print(result)
         assert len(result) in [0, 1]
@@ -148,7 +148,7 @@ class TestLibraryLevel:
                 "modality": "text",
             },
         ]
-        result = SecurityScanner.scan(llmSpec, maxBudget, datasets, max_th)
+        result = AgenticSecurity.scan(llmSpec, maxBudget, datasets, max_th)
         assert isinstance(result, dict)
         print(result)
         assert len(result) in [0, 1]
@@ -164,10 +164,10 @@ class TestEntrypointCI:
             temp_path = os.path.join(tmpdir, "custom_agesec.toml")
 
             # Override default_path to the temporary path
-            SecurityScanner.default_path = temp_path
+            AgenticSecurity.default_path = temp_path
 
             # Generate the default configuration
-            security = SecurityScanner()
+            security = AgenticSecurity()
             security.generate_default_settings()
 
             # Check that the config file was created at the temporary path
@@ -189,17 +189,17 @@ class TestEntrypointCI:
             temp_path = os.path.join(tmpdir, "custom_agesec.toml")
 
             # Override default_path to the temporary path
-            SecurityScanner.default_path = temp_path
+            AgenticSecurity.default_path = temp_path
 
             # Generate the default configuration
-            security = SecurityScanner()
+            security = AgenticSecurity()
             security.generate_default_settings()
 
             # Load the generated configuration
-            SecurityScanner.load_config(temp_path)
+            AgenticSecurity.load_config(temp_path)
 
             # Validate loaded configuration
-            config = SecurityScanner.config
+            config = AgenticSecurity.config
             assert (
                 config["general"]["maxBudget"] == 1000000
             ), "maxBudget should be 1000000"

codebeaver.yml

@@ -0,0 +1,2 @@
+from:python-pytest-poetry
+# This file was generated automatically by CodeBeaver based on your repository. Learn how to customize it here: https://docs.codebeaver.ai/configuration/

docs/http_spec.md

@@ -33,7 +33,7 @@ The `LLMSpec` class is the core of the HTTP specification. It provides the follo
 ### Methods
 
 - **`from_string(http_spec: str) -> LLMSpec`**: Parses an HTTP specification string into an `LLMSpec` object.
-- **`validate(prompt: str, encoded_image: str, encoded_audio: str, files: dict) -> null`**: Validates the request parameters based on the specified modality.
+- **`validate(prompt: str, encoded_image: str, encoded_audio: str, files: dict) -> None`**: Validates the request parameters based on the specified modality.
 - **`probe(prompt: str, encoded_image: str = "", encoded_audio: str = "", files: dict = {}) -> httpx.Response`**: Sends an HTTP request using the specified parameters.
 - **`verify() -> httpx.Response`**: Verifies the HTTP specification by sending a test request.
 
@@ -52,11 +52,12 @@ Authorization: Bearer sk-xxxxxxxxx
 Content-Type: application/json
 
 {
-  "model": "gpt-3.5-turbo",
-  "messages": [{"role": "user", "content": "<<PROMPT>>"}],
-  "temperature": 0.7
+    "model": "gpt-3.5-turbo",
+    "messages": [{"role": "user", "content": "<<PROMPT>>"}],
+    "temperature": 0.7
 }
 """
+
 spec = LLMSpec.from_string(http_spec)
 response = await spec.probe("What is the capital of France?")
 ```
@@ -70,11 +71,12 @@ Authorization: Bearer sk-xxxxxxxxx
 Content-Type: application/json
 
 {
-  "model": "gpt-4-vision-preview",
-  "messages": [{"role": "user", "content": "What is in this image? <<BASE64_IMAGE>>"}],
-  "temperature": 0.7
+    "model": "gpt-4-vision-preview",
+    "messages": [{"role": "user", "content": "What is in this image? <<BASE64_IMAGE>>"}],
+    "temperature": 0.7
 }
 """
+
 spec = LLMSpec.from_string(http_spec)
 encoded_image = encode_image_base64_by_url("https://example.com/image.jpg")
 response = await spec.probe("What is in this image?", encoded_image=encoded_image)
@@ -89,11 +91,12 @@ Authorization: Bearer sk-xxxxxxxxx
 Content-Type: application/json
 
 {
-  "model": "whisper-large-v3",
-  "messages": [{"role": "user", "content": "Transcribe this audio: <<BASE64_AUDIO>>"}],
-  "temperature": 0.7
+    "model": "whisper-large-v3",
+    "messages": [{"role": "user", "content": "Transcribe this audio: <<BASE64_AUDIO>>"}],
+    "temperature": 0.7
 }
 """
+
 spec = LLMSpec.from_string(http_spec)
 encoded_audio = encode_audio_base64_by_url("https://example.com/audio.mp3")
 response = await spec.probe("Transcribe this audio:", encoded_audio=encoded_audio)
@@ -108,11 +111,12 @@ Authorization: Bearer sk-xxxxxxxxx
 Content-Type: multipart/form-data
 
 {
-  "model": "gpt-3.5-turbo",
-  "messages": [{"role": "user", "content": "Process this file: <<FILE>>"}],
-  "temperature": 0.7
+    "model": "gpt-3.5-turbo",
+    "messages": [{"role": "user", "content": "Process this file: <<FILE>>"}],
+    "temperature": 0.7
 }
 """
+
 spec = LLMSpec.from_string(http_spec)
 files = {"file": ("document.txt", open("document.txt", "rb"))}
 response = await spec.probe("Process this file:", files=files)

docs/index.md

@@ -21,4 +21,4 @@ Note: Please be aware that Agentic Security is designed as a safety scanner tool
 
 ## UI 🧙
 
-<img width="100%" alt="booking-screen" src="https://res.cloudinary.com/dq0w2rtm9/image/upload/v1741192668/final_aa9jhb.gif">
+<img width="100%" alt="booking-screen" src="https://res.cloudinary.com/dq0w2rtm9/image/upload/v1736433557/z0bsyzhsqlgcr3w4ovwp.gif">

docs/probe_data.md

@@ -54,15 +54,20 @@ The `probe_data` module is a core component of the Agentic Security project, res
 
 - **Classes:**
   - `PromptSelectionInterface`: Abstract base class for prompt selection strategies.
+
     - Methods:
       - `select_next_prompt(current_prompt: str, passed_guard: bool) -> str`: Selects next prompt
       - `select_next_prompts(current_prompt: str, passed_guard: bool) -> list[str]`: Selects multiple prompts
-      - `update_rewards(previous_prompt: str, current_prompt: str, reward: float, passed_guard: bool) -> null`: Updates rewards
+      - `update_rewards(previous_prompt: str, current_prompt: str, reward: float, passed_guard: bool) -> None`: Updates rewards
+
   - `RandomPromptSelector`: Basic random selection with history tracking.
+
     - Parameters:
       - `prompts: list[str]`: List of available prompts
       - `history_size: int = 3`: Size of history to prevent cycles
+
   - `CloudRLPromptSelector`: Cloud-based RL implementation with fallback.
+
     - Parameters:
       - `prompts: list[str]`: List of available prompts
       - `api_url: str`: URL of RL service
@@ -70,7 +75,9 @@ The `probe_data` module is a core component of the Agentic Security project, res
       - `history_size: int = 300`: Size of history
       - `timeout: int = 5`: Request timeout
       - `run_id: str = ""`: Unique run identifier
+
   - `QLearningPromptSelector`: Local Q-learning implementation.
+
     - Parameters:
       - `prompts: list[str]`: List of available prompts
       - `learning_rate: float = 0.1`: Learning rate
@@ -79,11 +86,13 @@ The `probe_data` module is a core component of the Agentic Security project, res
       - `exploration_decay: float = 0.995`: Exploration decay rate
       - `min_exploration: float = 0.01`: Minimum exploration rate
       - `history_size: int = 300`: Size of history
-- **Module**: Main class that uses CloudRLPromptSelector.
-  - Parameters:
-    - `prompt_groups: list[str]`: Groups of prompts
-    - `tools_inbox: asyncio.Queue`: Queue for tool communication
-    - `opts: dict = {}`: Configuration options
+
+  - `Module`: Main class that uses CloudRLPromptSelector.
+
+    - Parameters:
+      - `prompt_groups: list[str]`: Groups of prompts
+      - `tools_inbox: asyncio.Queue`: Queue for tool communication
+      - `opts: dict = {}`: Configuration options
 
 ## Usage Examples
 
@@ -110,9 +119,10 @@ from agentic_security.probe_data.modules.rl_model import QLearningPromptSelector
 
 prompts = ["What is AI?", "Explain machine learning"]
 selector = QLearningPromptSelector(prompts)
+
 current_prompt = "What is AI?"
-next_prompt = selector.select_next_prompt(current_prompt, passed_guard=true)
-selector.update_rewards(current_prompt, next_prompt, reward=1.0, passed_guard=true)
+next_prompt = selector.select_next_prompt(current_prompt, passed_guard=True)
+selector.update_rewards(current_prompt, next_prompt, reward=1.0, passed_guard=True)
 ```
 
 ## Conclusion

docs/stylesheets/extra.css

@@ -1,16 +1,14 @@
 :root {
-  --md-primary-fg-color: #2E4053;
-  /* Primary color changed to pinkish */
-  --md-primary-fg-color--light: #E0A3B6;
-  --md-primary-fg-color--dark: #1C3F74;
-  /* Dark variant changed to blue */
+  --md-primary-fg-color: #e92063;
+  --md-primary-fg-color--light: #e92063;
+  --md-primary-fg-color--dark: #e92063;
 }
 
-/* Updated slate color scheme with new background */
+
+/* Revert hue value to that of pre mkdocs-material v9.4.0 */
 [data-md-color-scheme="slate"] {
   --md-hue: 230;
-  --md-default-bg-color: #1A1A1A;
-  /* Background changed to dark gray */
+  --md-default-bg-color: hsla(230, 15%, 21%, 1);
 }
 
 .hide {
@@ -26,15 +24,12 @@ img.index-header {
   max-width: 500px;
 }
 
-/* Updated custom colors */
 .pydantic-pink {
-  color: #E0A3B6;
-  /* Updated to match new theme */
+  color: #FF007F;
 }
 
 .team-blue {
-  color: #1C3F74;
-  /* Updated to match new theme */
+  color: #0072CE;
 }
 
 .secure-green {
@@ -72,6 +67,7 @@ img.index-header {
   text-align: center;
 }
 
+
 /* Hide the entire footer */
 .md-footer {
   display: none;

mkdocs.yml

@@ -89,7 +89,7 @@ theme:
         name: Switch to light mode
   icon:
     repo: fontawesome/brands/github
-  favicon: https://res.cloudinary.com/dq0w2rtm9/image/upload/v1741195421/favicon_kuz6xr.png
+  favicon: "https://res.cloudinary.com/dq0w2rtm9/image/upload/v1737555066/r17hrkre246doczwmvbv.png"
 
 extra:
   generator: false

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "agentic_security"
-version = "0.7.4"
+version = "0.5.1"
 description = "Agentic LLM vulnerability scanner"
 authors = ["Alexander Miasoiedov <msoedov@gmail.com>"]
 maintainers = ["Alexander Miasoiedov <msoedov@gmail.com>"]
@@ -52,10 +52,8 @@ sentry_sdk = "^2.22.0"
 orjson = "^3.10"
 pyfiglet = "^1.0.2"
 termcolor = "^2.4.0"
-mcp = "^1.4.1"
 
 # garak = { version = "*", optional = true }
-pytest-xdist = "3.6.1"
 
 
 [tool.poetry.group.dev.dependencies]
@@ -70,7 +68,7 @@ pytest-mock = "^3.14.0"
 black = ">=24.10,<26.0"
 mypy = "^1.12.0"
 pre-commit = "^4.0.1"
-huggingface-hub = ">=0.25.1,<0.30.0"
+huggingface-hub = ">=0.25.1,<0.29.0"
 
 # Docs
 mkdocs = ">=1.4.2"
@@ -88,13 +86,7 @@ build-backend = "poetry.core.masonry.api"
 
 
 [tool.pytest.ini_options]
-addopts = "--durations=5 -m 'not slow' -n 3"
+addopts = "--durations=5 -m 'not slow'"
 asyncio_mode = "auto"
 asyncio_default_fixture_loop_scope = "function"
 markers = "slow: marks tests as slow"
-
-[project]
-# MCP requires the following fields to be present in the pyproject.toml file
-name = "agentic_security"
-version = "1.0.0"
-requires-python = ">=3.11"

tests/conftest.py

@@ -1,17 +1,8 @@
 import os
 
 import pytest
-from cache_to_disk import delete_old_disk_caches
-
-from agentic_security.logutils import logger
 
 
 def pytest_runtest_setup(item):
     if "slow" in item.keywords and not os.getenv("RUN_SLOW_TESTS"):
         pytest.skip("Skipping slow test")
-
-
-@pytest.fixture(autouse=True, scope="session")
-def setup_delete_old_disk_caches():
-    logger.info("delete_old_disk_caches")
-    delete_old_disk_caches()

tests/test_app.py

@@ -0,0 +1,161 @@
+import pytest
+import asyncio
+from fastapi import FastAPI
+from asyncio import Queue, Event
+from agentic_security.core.app import create_app, get_tools_inbox, get_stop_event, get_current_run, set_current_run
+
+class TestApp:
+    """Test suite for agentic_security.core.app module."""
+
+    def test_create_app(self):
+        """Test that create_app returns a FastAPI instance."""
+        app = create_app()
+        assert isinstance(app, FastAPI)
+
+    @pytest.mark.asyncio
+    async def test_get_tools_inbox(self):
+        """Test that get_tools_inbox returns the global Queue instance."""
+        queue1 = get_tools_inbox()
+        await queue1.put("test item")
+        queue2 = get_tools_inbox()
+        result = queue2.get_nowait()
+        assert result == "test item"
+
+    def test_get_stop_event(self):
+        """Test that get_stop_event returns the global Event instance and is not set initially."""
+        event = get_stop_event()
+        assert isinstance(event, Event)
+        assert not event.is_set()
+
+    def test_current_run_initial(self):
+        """Test that get_current_run returns the global current_run with default values initially."""
+        run = get_current_run()
+        # Default values should be empty strings
+        assert run["spec"] == ""
+        assert run["id"] == ""
+
+    def test_set_current_run(self):
+        """Test that set_current_run correctly updates current_run."""
+        spec = "test run"
+        result = set_current_run(spec)
+        expected_id = hash(id(spec))
+        # Verify that spec is set correctly
+        assert result["spec"] == spec
+        assert result["id"] == expected_id
+
+    def test_current_run_after_set(self):
+        """Test that get_current_run returns the updated current_run after set_current_run is called."""
+        spec = "another test run"
+        set_current_run(spec)
+        current = get_current_run()
+        assert current["spec"] == spec
+        assert current["id"] == hash(id(spec))
+    def test_tools_inbox_same_instance(self):
+        """Test that get_tools_inbox returns the same Queue instance by default."""
+        queue1 = get_tools_inbox()
+        queue2 = get_tools_inbox()
+        assert queue1 is queue2
+
+    def test_stop_event_set(self):
+        """Test that setting the stop event is reflected in subsequent calls."""
+        event = get_stop_event()
+        event.set()  # set the global event
+        # Now, subsequent calls should return the same event which is set.
+        event2 = get_stop_event()
+        assert event2.is_set()
+
+    def test_set_current_run_with_none(self):
+        """Test that set_current_run handles None as a valid input and updates current_run accordingly."""
+        result = set_current_run(None)
+        expected_id = hash(id(None))
+        assert result["spec"] is None
+        assert result["id"] == expected_id
+
+    def test_multiple_current_run_assignments(self):
+        """Test multiple assignments to current_run to ensure it always updates correctly."""
+        first_spec = "first run"
+        result1 = set_current_run(first_spec)
+        expected_id1 = hash(id(first_spec))
+        assert result1["spec"] == first_spec
+        assert result1["id"] == expected_id1
+
+        second_spec = "second run"
+        result2 = set_current_run(second_spec)
+        expected_id2 = hash(id(second_spec))
+        assert result2["spec"] == second_spec
+        assert result2["id"] == expected_id2
+
+        current = get_current_run()
+        # The current_run should reflect the latest assignment.
+        assert current["spec"] == second_spec
+        assert current["id"] == expected_id2
+    @pytest.mark.asyncio
+    async def test_empty_tools_inbox_exception(self):
+        """Test that calling get_nowait on an empty tools_inbox raises QueueEmpty."""
+        from asyncio import QueueEmpty
+        queue = get_tools_inbox()
+        # Clear any existing items in the queue
+        while True:
+            try:
+                queue.get_nowait()
+            except QueueEmpty:
+                break
+        with pytest.raises(QueueEmpty):
+            queue.get_nowait()
+
+    def test_set_current_run_with_dict(self):
+        """Test that set_current_run correctly handles a dictionary input as spec."""
+        spec = {"key": "value"}
+        result = set_current_run(spec)
+        expected_id = hash(id(spec))
+        assert result["spec"] == spec
+        assert result["id"] == expected_id
+    @pytest.mark.asyncio
+    async def test_stop_event_wait(self):
+        """Test that waiting on the stop event returns once the event is set."""
+        event = get_stop_event()
+        event.clear()  # ensure event is not set
+        async def waiter():
+            await event.wait()
+            return True
+        waiter_task = asyncio.create_task(waiter())
+        # Wait a moment to ensure the waiter is pending
+        await asyncio.sleep(0.1)
+        assert not waiter_task.done()
+        event.set()
+        result = await waiter_task
+        assert result is True
+
+    def test_set_current_run_with_int(self):
+        """Test that set_current_run handles an integer input as spec."""
+        spec = 12345
+        result = set_current_run(spec)
+        expected_id = hash(id(spec))
+        assert result["spec"] == spec
+        assert result["id"] == expected_id
+
+    def test_create_app_routes(self):
+        """Test that create_app returns a FastAPI instance with default routes available."""
+        app = create_app()
+        paths = [route.path for route in app.routes]
+        # Check that the default OpenAPI route exists
+        assert "/openapi.json" in paths
+
+    @pytest.mark.asyncio
+    async def test_tools_inbox_async_put_get_order(self):
+        """Test that tools_inbox preserves order when items are added and retrieved asynchronously."""
+        queue = get_tools_inbox()
+        # Clear any existing items in the queue
+        from asyncio import QueueEmpty
+        while True:
+            try:
+                queue.get_nowait()
+            except QueueEmpty:
+                break
+        items = ["first", "second", "third"]
+        for item in items:
+            await queue.put(item)
+        result_items = []
+        for _ in items:
+            result_items.append(await queue.get())
+        assert result_items == items

tests/test_http_spec.py

@@ -0,0 +1,341 @@
+import pytest
+import base64
+import httpx
+import asyncio
+from agentic_security.http_spec import (
+    LLMSpec,
+    parse_http_spec,
+    escape_special_chars_for_json,
+    encode_image_base64_by_url,
+    encode_audio_base64_by_url,
+    InvalidHTTPSpecError,
+    Modality
+)
+
+################################################################################
+# Tests for agentic_security/http_spec.py
+################################################################################
+
+def test_escape_special_chars_for_json():
+    """Test escaping special characters in a prompt for JSON safety."""
+    prompt = 'Line1\nLine2\t"Quote"\\Backslash'
+    escaped = escape_special_chars_for_json(prompt)
+    assert '\\n' in escaped
+    assert '\\t' in escaped
+    assert '\\"' in escaped
+    assert '\\\\' in escaped
+
+def test_parse_http_spec_text():
+    """Test parsing a text HTTP spec without image/audio/files requirements."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\nThis is a prompt: <<PROMPT>>"
+    llm_spec = parse_http_spec(spec)
+    assert llm_spec.method == "POST"
+    assert llm_spec.url == "http://example.com/api"
+    assert llm_spec.headers["Content-Type"] == "application/json"
+    assert "<<PROMPT>>" in llm_spec.body
+    assert not llm_spec.has_files
+    assert not llm_spec.has_image
+    assert not llm_spec.has_audio
+
+def test_parse_http_spec_files():
+    """Test parsing a HTTP spec with multipart/form-data header indicating files."""
+    spec = "PUT http://example.com/upload\nContent-Type: multipart/form-data\n\nFile upload test"
+    llm_spec = parse_http_spec(spec)
+    assert llm_spec.has_files
+
+def test_parse_http_spec_image_audio():
+    """Test parsing a HTTP spec that requires image and audio via placeholders."""
+    spec = "GET http://example.com/api\nContent-Type: application/json\n\nImage: <<BASE64_IMAGE>> and Audio: <<BASE64_AUDIO>>"
+    llm_spec = parse_http_spec(spec)
+    assert llm_spec.has_image
+    assert llm_spec.has_audio
+
+def test_encode_image_base64_by_url(monkeypatch):
+    """Test that image encoding returns the correct base64 string with prefix."""
+    dummy_content = b'test_image'
+    class DummyResponse:
+        def __init__(self, content):
+            self.content = content
+
+    def dummy_get(url):
+        return DummyResponse(dummy_content)
+
+    monkeypatch.setattr(httpx, "get", dummy_get)
+    result = encode_image_base64_by_url("http://dummyurl.com/image.jpg")
+    expected = "data:image/jpeg;base64," + base64.b64encode(dummy_content).decode("utf-8")
+    assert result == expected
+
+def test_encode_audio_base64_by_url(monkeypatch):
+    """Test that audio encoding returns the correct base64 string with prefix."""
+    dummy_content = b'test_audio'
+    class DummyResponse:
+        def __init__(self, content):
+            self.content = content
+
+    def dummy_get(url):
+        return DummyResponse(dummy_content)
+
+    monkeypatch.setattr(httpx, "get", dummy_get)
+    result = encode_audio_base64_by_url("http://dummyurl.com/audio.mp3")
+    expected = "data:audio/mpeg;base64," + base64.b64encode(dummy_content).decode("utf-8")
+    assert result == expected
+
+@pytest.mark.asyncio
+async def test_probe_text(monkeypatch):
+    """Test the probe function for text modality by replacing <<PROMPT>>."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\n{\"prompt\": \"<<PROMPT>>\"}"
+    llm_spec = parse_http_spec(spec)
+
+    async def dummy_request(self, method, url, headers, content, timeout):
+        return httpx.Response(200, text="ok")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    response = await llm_spec.probe("Hello")
+    assert response.status_code == 200
+    assert "ok" in response.text
+
+@pytest.mark.asyncio
+async def test_probe_with_files(monkeypatch):
+    """Test that probe correctly branches to _probe_with_files when files are provided."""
+    spec = "POST http://example.com/api\nContent-Type: multipart/form-data\n\nFile data"
+    llm_spec = parse_http_spec(spec)
+    files = {"file": ("dummy.txt", b"data")}
+
+    async def dummy_request(self, method, url, headers, files, timeout):
+        return httpx.Response(200, text="file upload ok")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    response = await llm_spec.probe("Unused", files=files)
+    assert response.status_code == 200
+    assert "file upload ok" in response.text
+
+@pytest.mark.asyncio
+async def test_verify_image(monkeypatch):
+    """Test verify method branch for image modality by monkeypatching image encoder."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\n{\"image\": \"<<BASE64_IMAGE>>\"}"
+    llm_spec = parse_http_spec(spec)
+
+    # Replace the image encoder to return a dummy string
+    monkeypatch.setattr("agentic_security.http_spec.encode_image_base64_by_url", lambda url="": "dummy_image")
+
+    async def dummy_request(self, method, url, headers, content, timeout):
+        # Check that the dummy image is injected in the content
+        assert "dummy_image" in content
+        return httpx.Response(200, text="image ok")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    response = await llm_spec.verify()
+    assert response.status_code == 200
+    assert "image ok" in response.text
+
+@pytest.mark.asyncio
+async def test_verify_audio(monkeypatch):
+    """Test verify method branch for audio modality by monkeypatching audio encoder."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\n{\"audio\": \"<<BASE64_AUDIO>>\"}"
+    llm_spec = parse_http_spec(spec)
+
+    monkeypatch.setattr("agentic_security.http_spec.encode_audio_base64_by_url", lambda url: "dummy_audio")
+
+    async def dummy_request(self, method, url, headers, content, timeout):
+        # Ensure that the dummy audio string is present in the request content
+        assert "dummy_audio" in content
+        return httpx.Response(200, text="audio ok")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    response = await llm_spec.verify()
+    assert response.status_code == 200
+    assert "audio ok" in response.text
+
+@pytest.mark.asyncio
+async def test_verify_files(monkeypatch):
+    """Test verify method branch for files modality where _probe_with_files is invoked."""
+    spec = "POST http://example.com/api\nContent-Type: multipart/form-data\n\nFile data"
+    llm_spec = parse_http_spec(spec)
+
+    async def dummy_request(self, method, url, headers, files, timeout):
+        return httpx.Response(200, text="files ok")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    response = await llm_spec.verify()
+    assert response.status_code == 200
+    assert "files ok" in response.text
+
+def test_llm_spec_modality_property():
+    """Test that the modality property reflects the correct modality."""
+    spec_text = "POST http://example.com/api\nContent-Type: application/json\n\nPrompt: <<PROMPT>>"
+    llm_spec_text = parse_http_spec(spec_text)
+    assert llm_spec_text.modality == Modality.TEXT
+
+    spec_image = "POST http://example.com/api\nContent-Type: application/json\n\nImage: <<BASE64_IMAGE>>"
+    llm_spec_image = parse_http_spec(spec_image)
+    assert llm_spec_image.modality == Modality.IMAGE
+
+    spec_audio = "POST http://example.com/api\nContent-Type: application/json\n\nAudio: <<BASE64_AUDIO>>"
+    llm_spec_audio = parse_http_spec(spec_audio)
+    assert llm_spec_audio.modality == Modality.AUDIO
+
+def test_from_string_invalid():
+    """Test that LLMSpec.from_string raises an error for an invalid spec."""
+    invalid_spec = "INVALID_SPEC"
+    with pytest.raises(InvalidHTTPSpecError):
+        LLMSpec.from_string(invalid_spec)
+@pytest.mark.asyncio
+async def test_validate_missing_files():
+    """Test that LLMSpec.validate raises a ValueError when files are required but missing."""
+    spec = "POST http://example.com/api\nContent-Type: multipart/form-data\n\nFile upload test"
+    llm_spec = parse_http_spec(spec)
+    with pytest.raises(ValueError, match="Files are required"):
+        llm_spec.validate("test prompt", "", "", {})
+
+@pytest.mark.asyncio
+async def test_validate_missing_image():
+    """Test that LLMSpec.validate raises a ValueError when an image is required but missing."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\nImage: <<BASE64_IMAGE>>"
+    llm_spec = parse_http_spec(spec)
+    with pytest.raises(ValueError, match="An image is required"):
+        llm_spec.validate("test prompt", "", "dummy_audio", {})
+
+@pytest.mark.asyncio
+async def test_validate_missing_audio():
+    """Test that LLMSpec.validate raises a ValueError when audio is required but missing."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\nAudio: <<BASE64_AUDIO>>"
+    llm_spec = parse_http_spec(spec)
+    with pytest.raises(ValueError, match="Audio is required"):
+        llm_spec.validate("test prompt", "dummy_image", "", {})
+
+def test_fn_alias(monkeypatch):
+    """Test that LLMSpec.fn is a functional alias for LLMSpec.probe."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\n{\"prompt\": \"<<PROMPT>>\"}"
+    llm_spec = parse_http_spec(spec)
+
+    # Instead of overriding the instance method, verify the alias at the class level.
+    assert LLMSpec.fn is LLMSpec.probe
+
+def test_escape_special_chars_no_special():
+    """Test that the escape function returns the original string if no special characters are present."""
+    prompt = "Simple text without specials"
+    escaped = escape_special_chars_for_json(prompt)
+    assert escaped == "Simple text without specials"
+@pytest.mark.asyncio
+async def test_probe_text_with_special_chars(monkeypatch):
+    """Test probe for text modality with special characters in prompt ensuring escaped content."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\n{\"prompt\": \"<<PROMPT>>\"}"
+    llm_spec = parse_http_spec(spec)
+    captured = {}
+
+    async def dummy_request(self, method, url, headers, content, timeout):
+        captured['content'] = content
+        return httpx.Response(200, text="ok")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    test_prompt = 'Hello\nWorld\t"Test"'
+    response = await llm_spec.probe(test_prompt)
+    expected_escaped = escape_special_chars_for_json(test_prompt)
+    assert expected_escaped in captured['content']
+    assert response.status_code == 200
+
+@pytest.mark.asyncio
+async def test_verify_both_image_audio(monkeypatch):
+    """Test verify method when both image and audio placeholders are present.
+    Expect a ValueError because only the image branch is triggered by pattern matching and the missing audio causes validation to fail."""
+    spec = ("POST http://example.com/api\nContent-Type: application/json\n\n"
+            "{\"audio\": \"<<BASE64_AUDIO>>\", \"image\":\"<<BASE64_IMAGE>>\"}")
+    llm_spec = parse_http_spec(spec)
+    # Monkey patch the image encoder to return a dummy value
+    monkeypatch.setattr("agentic_security.http_spec.encode_image_base64_by_url", lambda url="": "dummy_image")
+    with pytest.raises(ValueError, match="Audio is required"):
+        await llm_spec.verify()
+
+def test_parse_http_spec_invalid_header_format():
+    """Test that parse_http_spec raises an error when a header line doesn't have the expected 'key: value' format."""
+    invalid_spec = "GET http://example.com/api\nInvalidHeaderWithoutColon\n\nBody with <<PROMPT>>"
+    with pytest.raises(ValueError):
+            parse_http_spec(invalid_spec)
+
+def test_from_string_valid():
+    """Test that LLMSpec.from_string returns a valid LLMSpec object when given a proper spec string."""
+    spec = "GET http://example.com/api\nContent-Type: application/json\n\n{ \"prompt\": \"<<PROMPT>>\" }"
+    llm_spec = LLMSpec.from_string(spec)
+    assert llm_spec.method == "GET"
+    assert llm_spec.url == "http://example.com/api"
+
+@pytest.mark.asyncio
+async def test_parse_http_spec_multiline_body():
+    """Test parsing an HTTP spec with a multiline body to ensure body concatenation works."""
+    spec = (
+        "PATCH http://example.com/api\n"
+        "Content-Type: application/json\n"
+        "\n"
+        "Line one of body\n"
+        "Line two of body\n"
+        "Line three"
+    )
+    llm_spec = parse_http_spec(spec)
+    # As implemented, the parser concatenates lines without newline delimiters
+    expected_body = "Line one of bodyLine two of bodyLine three"
+    assert llm_spec.body == expected_body
+
+@pytest.mark.asyncio
+async def test_encode_image_default_argument(monkeypatch):
+    """Test that encode_image_base64_by_url works with its default URL argument."""
+    dummy_content = b'default_image'
+    class DummyResponse:
+        def __init__(self, content):
+            self.content = content
+
+    def dummy_get(url):
+        # check that the default URL (which includes 'fluidicon.png') is used
+        assert "fluidicon.png" in url
+        return DummyResponse(dummy_content)
+
+    monkeypatch.setattr(httpx, "get", dummy_get)
+    result = encode_image_base64_by_url()
+    expected = "data:image/jpeg;base64," + base64.b64encode(dummy_content).decode("utf-8")
+    assert result == expected
+
+@pytest.mark.asyncio
+async def test_probe_without_prompt_placeholder(monkeypatch):
+    """Test the probe function when the request body does not include the <<PROMPT>> placeholder."""
+    spec = "POST http://example.com/api\nContent-Type: application/json\n\n{\"message\": \"No placeholder here\"}"
+    llm_spec = parse_http_spec(spec)
+
+    captured = {}
+
+    async def dummy_request(self, method, url, headers, content, timeout):
+        captured['content'] = content
+        return httpx.Response(200, text="ok without placeholder")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    response = await llm_spec.probe("Ignored prompt")
+    assert "No placeholder here" in captured['content']
+    assert response.status_code == 200
+
+def test_validate_success():
+    """Test that LLMSpec.validate does not raise an error when all required data is provided."""
+    # Test case for files: files are provided as required
+    spec_files = "POST http://example.com/api\nContent-Type: multipart/form-data\n\nFile upload"
+    llm_spec_files = parse_http_spec(spec_files)
+    llm_spec_files.validate("some prompt", "dummy_image", "dummy_audio", {"file": ("dummy.txt", b"data")})
+
+    # Test case for image: image is provided as required
+    spec_image = "POST http://example.com/api\nContent-Type: application/json\n\nImage: <<BASE64_IMAGE>>"
+    llm_spec_image = parse_http_spec(spec_image)
+    llm_spec_image.validate("some prompt", "dummy_image", "dummy_audio", {})
+
+    # Test case for audio: audio is provided as required
+    spec_audio = "POST http://example.com/api\nContent-Type: application/json\n\nAudio: <<BASE64_AUDIO>>"
+    llm_spec_audio = parse_http_spec(spec_audio)
+    llm_spec_audio.validate("some prompt", "dummy_image", "dummy_audio", {})
+
+@pytest.mark.asyncio
+async def test_probe_invalid_url(monkeypatch):
+    """Test that probe raises an exception when the HTTP client fails due to an invalid URL."""
+    spec = "GET http://nonexistent_url/api\nContent-Type: application/json\n\n{\"prompt\": \"<<PROMPT>>\"}"
+    llm_spec = parse_http_spec(spec)
+
+    async def dummy_request(self, method, url, headers, content, timeout):
+        raise httpx.RequestError("Invalid URL")
+
+    monkeypatch.setattr(httpx.AsyncClient, "request", dummy_request)
+    with pytest.raises(httpx.RequestError):
+        await llm_spec.probe("Test")

tests/test_mcp.py

@@ -1,12 +0,0 @@
-import pytest
-
-from agentic_security.mcp.client import run
-
-
-@pytest.mark.asyncio
-async def test_mcp_echo_tool():
-    """Test the echo tool functionality"""
-    prompts, resources, tools = await run()
-    assert prompts
-    assert resources
-    assert tools

ui/package-lock.json

@@ -6891,9 +6891,9 @@
       }
     },
     "node_modules/http-proxy-middleware": {
-      "version": "2.0.9",
-      "resolved": "https://registry.npmjs.org/http-proxy-middleware/-/http-proxy-middleware-2.0.9.tgz",
-      "integrity": "sha512-c1IyJYLYppU574+YI7R4QyX2ystMtVXZwIdzazUIPIJsHuWNd+mho2j+bKoHftndicGj9yh+xjd+l0yj7VeT1Q==",
+      "version": "2.0.7",
+      "resolved": "https://registry.npmjs.org/http-proxy-middleware/-/http-proxy-middleware-2.0.7.tgz",
+      "integrity": "sha512-fgVY8AV7qU7z/MmXJ/rxwbrtQH4jBQ9m7kp3llF0liB7glmFeVZFBepQb32T3y8n8k2+AEYuMPCpinYW+/CuRA==",
       "dev": true,
       "license": "MIT",
       "dependencies": {