agentic_security/docs/optimizer.md

Bayesian Optimization in Security Fuzzing

The fuzzer implements an optimization system using scikit-optimize (skopt) to minimize failure rates during security scans. This document explains the optimizer's implementation and behavior.

Overview

The optimizer is used in both single-shot and many-shot scanning modes when the optimize parameter is True. It dynamically adjusts scan parameters to minimize failure rates while staying within budget constraints.

Implementation Details

Initialization

The optimizer is initialized with:

Optimizer(
    [Real(0, 1)],  # Single parameter space (0 to 1)
    base_estimator="GP",  # Gaussian Process estimator
    n_initial_points=25  # Initial exploration points
)

Optimization Process

1. Parameter Space: A single real-valued parameter between 0 and 1
2. Objective: Minimize the failure rate (negative failure rate is maximized)
3. Update Mechanism:

   next_point = optimizer.ask()
   optimizer.tell(next_point, -failure_rate)
   

4. Early Stopping: If best failure rate exceeds 50%:

   if best_failure_rate > 0.5:
       yield ScanResult.status_msg(
           f"High failure rate detected ({best_failure_rate:.2%}). Stopping this module..."
       )
       break
   

Usage in Scanning

The optimizer is integrated into both scan types:

Single-shot Scan

• Used in perform_single_shot_scan()
• Optimizes failure rates across prompt modules
• Considers token budget constraints

Many-shot Scan

• Used in perform_many_shot_scan()
• Handles more complex multi-step attacks
• Maintains separate failure rate tracking

Key Parameters

Parameter	Description
base_estimator	Gaussian Process (GP) used for optimization
n_initial_points	25 initial exploration points
Real(0, 1)	Single parameter space being optimized
failure_rate	Current failure rate being minimized

Optimization Flow

1. Initialize optimizer with GP estimator
2. Collect initial 25 data points
3. For each prompt:
   • Calculate current failure rate
   • Update optimizer with new point
   • Check for early stopping conditions
4. Continue until scan completes or budget exhausted

Error Handling

The optimizer is wrapped in try/except blocks to ensure scan failures don't crash the entire process. Any optimization errors are logged and the scan continues with default parameters.