ASL V6

ASL V6 is an open-source AI red-teaming engine built for security researchers and developers working with LLM agents (LangChain, LangGraph, AutoGPT, etc.).

The problem: Standard security tools use regex and pattern matching on AI agent code. They flag every docstring, mock string, and unit test as a vulnerability because they don't understand code context. This creates hundreds of false positives that waste engineering time.

What ASL V6 does differently:

• Parses Python Abstract Syntax Trees (ASTs) instead of regex — ignores tests, comments, and docstrings automatically.

• Boots isolated Docker micro-containers to verify if a potential exploit actually executes at runtime.

• Connects to local Ollama for offline remediation code generation.

• Covers OWASP LLM Top 10 and MITRE ATLAS attack patterns.

Benchmark results: Tested on a 513-file LangGraph repository. Filtered 593 false alarms to isolate 13 verified security flaws.

Built by a 22-year-old founder while researching 6+ CVEs in AI infrastructure, including vulnerabilities in AutoGPT, FlowiseAI, and LLaMmlein. Acknowledged by BSI Germany, CERT-EE Estonia, and NIIS (Estonia).

MIT License. Free to use, fork, and contribute.