This is a multi-agent framework designed to automate the end-to-end process of reproducing security vulnerabilities. It leverages a coordinated system of specialized agents to handle different aspects of the reproduction pipeline.

SecVerifier employs a hierarchical multi-agent architecture:

1. ManagerAgent - The coordinator agent that orchestrates the full process
2. BuilderAgent - Specialized assistant responsible for fixing and optimizing build processes for code repositories
3. ExploiterAgent - Specialized assistant responsible for crafting proof-of-concept (PoC) exploits to reproduce vulnerabilities in code repositories
4. FixerAgent - Specialized assistant responsible for creating a unified patch file that fixes vulnerabilities in code repositories.

The system uses agent delegation to pass control between specialized agents while maintaining runtime state to preserve the container environment integrity.

• Coordinated Multi-Agent Workflow: Specialized agents work together in sequence
• Maintained Runtime State: Container state is preserved across agent transitions
• Vulnerability-Specific Inputs: Each agent receives vulnerability-specific information
• Comprehensive Outputs: Detailed outputs from each phase of reproduction

# Clone the repository
git clone https://github.com/SEC-bench/SecVerifier.git
cd SecVerifier

# Install dependencies with Poetry
poetry install

Usage: ./run_multi-agent.sh -m <max_iterations> [-l <llm_config>] [-c <condenser>] [-i <instance_id>] [-s <limit>] [-t <max_budget>] [-b <label>]
  -m: Maximum iterations (default: 20)
  -l: LLM config (default: llm.4o)
  -c: Condenser (default: recent)
  -i: Instance ID
  -s: Limit the number of instances to process
  -t: Maximum budget per task
  -b: Label
  -w: Number of workers (default: 1)
  -h: Show this help message

1. Initialization: The system loads vulnerability data from SEC-bench
2. Builder Phase: BuilderAgent validates and fixes build scripts
3. Exploiter Phase: ExploiterAgent improves helper scripts and creates PoC files
4. Fixer Phase: FixerAgent creates a unified patch file that fixes vulnerabilities in code repositories
5. Evaluation: The system evaluates the success of the reproduction process