Dynamic penetration testing and excessive agency (OWASP LLM08) auditing toolkit for multi-hop AI agents across LangChain, CrewAI, AutoGen, and Model Context Protocol (MCP) ecosystems.
The OWASP Agentic Threat Auditor (OATA) is an open-source dynamic penetration testing and red teaming toolkit designed to audit autonomous AI agents for Excessive Agency (OWASP LLM08), parameter mutation escapes, and Layer 7 lateral delegation escalation.
Static code analysis tools only inspect syntax. OATA dynamically interacts with agent runtime middleware, executing mutated cognitive payloads against live tool-calling interfaces to verify whether execution boundaries remain isolated from untrusted prompts.
Key Capabilities:
Dynamic Middleware Fuzzing: Probes live tool dispatch loops for parameter injection, format string confusion, and path traversal mutations.
Safe Local Containment: Defaults strictly to loopback targets (127.0.0.1, localhost, and local sandboxes) to prevent unauthorized remote scanning.
Pluggable Framework Adapters: Native drivers for LangChain, CrewAI, AutoGen, and Model Context Protocol (MCP) servers.
Standardized Machine Reporting: Generates terminal vulnerability summaries, JSON dumps, and native SARIF v2.1.0 output for direct GitHub Code Scanning integration.
Offline Trace Replay: Audits pre-recorded agent execution traces without consuming live model API tokens.
Target Threat Taxonomy:
OATA-VEC-01 (Parameter Mutation Escape): Tests tool arguments for path traversal and schema boundary escapes.
OATA-VEC-02 (Unbounded Scope Escalation): Probes runtime authorization when prompts demand privileged execution shells.
OATA-VEC-03 (Context Boundary Bleed): Probes multi-turn conversational context for prompt injection overrides and lateral credential queries.
Quickstart:
Clone the repository from GitHub: git clone https://github.com/OWASP/Agentic-Threat-Auditor.git
Install in editable mode: pip install -e .
Run the local demo: oata demo