# NeuroSploit v3.4.0

![NeuroSploit](https://img.shields.io/badge/NeuroSploit-Autonomous%20AI%20Pentest-blueviolet)
![Version](https://img.shields.io/badge/Version-3.4.0-blue)
![License](https://img.shields.io/badge/License-MIT-green)
![Harness](https://img.shields.io/badge/Harness-Rust%20%7C%20tokio%20%7C%20axum-e6b673)
![Agents](https://img.shields.io/badge/MD%20Agents-249-red)
![Models](https://img.shields.io/badge/Models-12%20providers%20%2F%2040%2B-success)
![Backends](https://img.shields.io/badge/Subscription-Claude%20%7C%20Codex%20%7C%20Grok%20%7C%20Gemini-informational)
![MCP](https://img.shields.io/badge/MCP-Playwright-orange)

**Autonomous, markdown-driven AI penetration testing — now with a Rust multi-model harness.**

NeuroSploit turns a URL (or a code repository) into an autonomous security
engagement. A high-performance **Rust harness** (`tokio` + `axum`) drives a
**pool of LLM models** with concurrency, **provider failover**, and **N-model
validator voting** — multiple models must independently agree a finding is real
before it is reported. After recon, the harness **intelligently selects** which
of the **249 markdown agents** match the target instead of running them blindly,
learns across runs via a **reinforcement-learning** reward loop, and serves its
own polished web dashboard.

> The Python engine (v3.3.0) and the original monolith live in
> [`legacy/`](legacy/README.md); the v3.3.0 stdlib dashboard remains in `webgui/`.

## 🦀 The Rust harness (`neurosploit-rs/`)

```bash
cd neurosploit-rs && cargo build --release

# Web dashboard (black-box + white-box modes)
./target/release/neurosploit serve                       # → http://127.0.0.1:8788

# Black-box: recon → intelligent agent selection → parallel exploit → vote → report
./target/release/neurosploit run https://target.example \
    --model anthropic:claude-opus-4-8 --model openai:gpt-5.1 --vote-n 3

# White-box: analyse a repository's source for vulnerabilities
./target/release/neurosploit whitebox /path/to/repo --subscription --model anthropic:claude-opus-4-8

# Subscription (no API key) + real browser proof via Playwright MCP
./target/release/neurosploit run https://t.example --subscription --mcp --model anthropic:claude-opus-4-8

# Pipeline self-test, no keys/login required
./target/release/neurosploit run https://t.example --offline
```

**What it does**

- **Two modes** — *black-box* (URL recon → exploit) and *white-box* (walk a repo,
  run code-review/SAST agents on the source).
- **Intelligent selection** — the model picks the agents whose preconditions match
  the recon, then runs that subset (not top-N).
- **Multi-model pool** — bounded concurrency, **provider failover**, and the same
  panel forms the **N-model validator jury** that cuts false positives.
- **Two auth paths** — **model APIs** (provider key) *or* **subscription**: drive
  your local **Claude Code / Codex / Grok / Gemini** logins directly, no API key.
- **12 providers / 40+ models** (Claude, GPT, Grok, **Gemini**, NVIDIA NIM,
  DeepSeek, Mistral, Qwen, Groq, Together, OpenRouter, Ollama).
- **RL rewards** persisted to `data/rl_state_rs.json` — validated findings reward
  an agent, biasing the next run.
- **Artifacts for reuse** — every run writes `runs/<target>-<ts>/`:
  `recon.json/md`, `exploitation.md`, `findings.json/md`, `report.html`.
- **Playwright MCP** on the subscription path for real browser-based proof.

### Agent library — 249 agents

| Category | Dir | Count | Purpose |
|----------|-----|-------|---------|
| Vulnerability specialists | `agents_md/vulns/` | 196 | Exploit a specific vuln class |
| Recon | `agents_md/recon/` | 12 | Information gathering / attack surface |
| Code (white-box SAST) | `agents_md/code/` | 24 | Source-code vulnerability review |
| Meta | `agents_md/meta/` | 17 | Orchestrator, validator, scorers, reporter, RL |

---

## Why this architecture

| Old (≤ v3.2.4) | New (v3.3.0) |
|----------------|-------------|
| 2,500-line Python orchestrator + hand-coded agent classes | Markdown agents + thin engine |
| One embedded LLM loop | Pluggable agentic CLI backends (Claude/Codex/Grok) |
| Provider SDK juggling | Backend owns the agent loop; engine just composes & collects |
| Static agent list | RL-weighted, recon-aware agent selection |
| Reflection-based "evidence" | Playwright MCP proof-of-execution + adversarial validation |

---

## How it works

```
          ┌──────────────────────────────────────────────────────────────┐
   URL ──▶ │  neurosploit (terminal)                                       │
          │     │                                                          │
          │     ▼                                                          │
          │  orchestrator ── loads agents_md/ (213) ── applies RL weights  │
          │     │                                                          │
          │     ▼  composes ONE master prompt                              │
          │  backend (Claude Code | Codex | Grok)  ◀── Playwright MCP      │
          │     │  autonomously runs the pipeline below                    │
          │     ▼                                                          │
          │  recon → select agents → exploit → VALIDATE → filter FPs       │
          │        → severity → impact → report → RL feedback              │
          └──────────────────────────────────────────────────────────────┘
                       │                          │
                       ▼                          ▼
              results/findings.json        data/rl_state.json (learns)
```

The engine never fabricates findings: every candidate is independently
re-exploited (`meta/exploit_validator`), run through an adversarial skeptic
(`meta/false_positive_filter`), and only then scored and reported.

---

## The agent library (`agents_md/`)

**213 agents** — see [`agents_md/REGISTRY.md`](agents_md/REGISTRY.md).

- **196 vulnerability specialists** (`agents_md/vulns/`) — each a self-contained
  playbook with a real methodology, payloads, CWE mapping, and a strict
  anti-false-positive `## System Prompt`. Coverage includes the classic OWASP
  web set **plus modern classes**:
  - **LLM/AI security** (OWASP LLM Top 10): prompt injection (direct/indirect),
    jailbreak, system-prompt leak, insecure output handling, RAG poisoning,
    tool-invocation/function-calling abuse, excessive agency, PII leakage…
  - **Cloud/K8s/containers**: IMDS SSRF (AWS/GCP/Azure), kubelet/dashboard
    exposure, container & docker-socket escape, bucket takeover, IAM privesc…
  - **Modern API/auth**: JWT alg/kid/jwk confusion, OAuth PKCE downgrade, SAML
    XSW, OIDC, CSWSH, refresh-token & MFA bypass, account-takeover chains…
  - **Advanced injection**: SSTI (Jinja2/FreeMarker/Velocity/Thymeleaf), SSPP,
    XXE OOB, YAML/pickle deserialization, JNDI, XSLT…
  - **Protocol/cache/smuggling**: HTTP/2 & CL.TE/TE.CL desync, h2c, web cache
    deception/poisoning, response splitting, path-confusion…
  - **Logic/crypto/supply-chain**: dependency confusion, padding oracle, weak
    JWT secret, price/coupon/workflow abuse, exposed `.git`/`.env`/CI secrets…

- **17 meta-agents** (`agents_md/meta/`): `orchestrator`, `recon`,
  `exploit_validator`, `false_positive_filter`, `severity_assessor`,
  `impact_evaluator`, `reporter`, `rl_feedback`, plus migrated expert roles.

Add your own by dropping a `.md` into `agents_md/vulns/` (or extend the
data-driven builder, `scripts/build_agents.py`). It is picked up automatically.

---

## Quickstart

```bash
# 1. Have at least one agentic CLI installed: Claude Code, Codex, or Grok CLI
#    (Playwright MCP needs Node/npx)
./neurosploit backends          # show what's detected
./neurosploit agents            # {'vulns': 196, 'meta': 17, 'total': 213}

# 2. Interactive: enter a URL, pick a backend + model, go
./neurosploit

# 3. Or one-shot:
./neurosploit run https://target.example \
    --backend claude --model claude-opus-4-8 \
    --collaborator oob.your-collab.net

# 4. Preview the composed master prompt without executing the backend:
./neurosploit run https://target.example --dry-run
```

Outputs land in `results/<target>/findings.json` and `reports/`, and the RL
state updates in `data/rl_state.json`.

### Web dashboard

A zero-dependency (Python stdlib only) dashboard — no npm, no build step:

```bash
python3 webgui/server.py        # → http://127.0.0.1:8787
```

Tabs:
- **Run** — multi-target input, backend + provider + model pickers (40 models
  across CLI and API providers), verbosity, RL/MCP toggles, a live execution
  console (shows the exact backend command and per-task activity), and findings
  with screenshots.
- **Agents** — browse all 213 agents and **add new `.md` agents** from the UI;
  the main orchestrator picks them up on the next run.
- **Insights** — interactive chart of RL agent weights + findings by severity.
- **Reports** — download/preview the **PDF + HTML** reports (Typst engine).
- **Settings · API** — execution mode (CLI vs API), per-provider API keys,
  orchestrator selection, default verbosity.

It calls `neurosploit_agent` directly. The previous React app and FastAPI backend
were retired to `legacy/` (`frontend_react/`, `backend_fastapi/`).

### Backends

| Backend | Binary | Autonomy flag | Subscription |
|---------|--------|---------------|--------------|
| Claude Code | `claude` | `--dangerously-skip-permissions` | ✅ via Claude login |
| Codex CLI | `codex` | `--dangerously-bypass-approvals-and-sandbox` | — |
| Grok CLI | `grok` | `--yolo` | — |

The engine auto-detects installed backends and only offers those. In the
interactive flow, answering **yes** to "Use Claude subscription" runs Claude Code
against your logged-in subscription instead of an API key.

### Models

Latest models per provider live in `neurosploit_agent/models.py`, including the
**NVIDIA NIM** provider (PR #28, OpenAI-compatible at
`https://integrate.api.nvidia.com/v1`, `nvapi-` keys), Anthropic Claude 4.x,
OpenAI, xAI Grok, Gemini, OpenRouter, and local Ollama.

---

## Reinforcement learning

Every run produces per-agent reward signals (`meta/rl_feedback` +
`neurosploit_agent/rl.py`): validated findings reward an agent (weighted by
severity), rejected false positives penalize it, correct skips stay neutral.
Weights are bounded `[0.05, 1.0]` and carry per-tech-stack affinity, so the
engine learns, e.g., to prioritize `ssti_jinja2` on Flask targets. State is
explainable and persisted to `data/rl_state.json`.

---

## Safety & authorization

NeuroSploit is for **authorized** security testing only. Every agent's system
prompt enforces scope and proof-of-exploitation; DoS-class agents refuse to
flood and require explicit rules-of-engagement. You are responsible for having
written permission for any target you point it at.

---

## Repository layout

```
neurosploit                 # launcher (./neurosploit)
neurosploit_agent/          # the v3.3.0 engine
  cli.py  orchestrator.py  agent_loader.py  backends.py  rl.py  mcp.py  models.py  config.py
agents_md/
  vulns/   (196)            # vulnerability specialist agents
  meta/    (17)             # orchestrator, recon, validator, scorers, reporter, RL, roles
  REGISTRY.md               # generated index
scripts/build_agents.py     # data-driven agent builder
legacy/                     # retired pre-v3.3.0 Python orchestration
```

See [`RELEASE.md`](RELEASE.md) for the full v3.3.0 changelog.

---

## License

MIT.