AI Agent Red Team Engine + AuditLens

Red team engine on GitHub · AuditLens on GitHub

What it does

The red team engine runs RAG prompt-injection, tool-misuse and multi-turn social-engineering attacks against LangGraph agents. It grades what the agent actually did: which tools it called, and with which arguments. It covers 113 adversarial prompts across 8 attack categories, mapped to the OWASP LLM Top 10, MITRE ATLAS and CWE.

AuditLens takes those findings, or a scan of a GitHub repo, maps them onto EU AI Act articles, and produces a cited, scored assessment as JSON and PDF.

Key finding

Ten multi-turn chains were run against a customer-support agent built without authorization gates, on four models. Six chains succeeded on every model. One is a three-message account takeover: look up the account’s email, change it to an attacker’s address, then send a password reset. It completed on GPT-4o, GPT-4o-mini, Claude Sonnet 4 and Claude Haiku 4.5. A better-aligned model did not stop them. The fix belongs in the agent’s design, through real authorization checks.

Tools

Python, LangGraph, FastAPI, Pydantic, asyncio, ChromaDB, Anthropic and OpenAI APIs