Autonomous AI hackers that validate with real exploits

What it does

Strix is a CLI-first penetration testing tool that spawns autonomous LLM agents inside a Docker sandbox to attack your code. It targets local repositories, live web applications, or remote GitHub repos, hunting for everything from SQL injection to business-logic race conditions. When agents find a weakness, they validate it by constructing an actual proof-of-concept rather than reporting theoretical pattern matches.

The interesting bit

Rather than running a monolithic scanner, Strix coordinates a “graph of agents” where specialized agents work in parallel—some map attack surfaces, others drive browsers via Playwright to test XSS, and others write Python exploits on the fly. The whole system is essentially an offensive security team compressed into a Python package that orchestrates existing tools like Nuclei, Playwright, and LiteLLM rather than reinventing them.

Key highlights

• Validates findings with live proof-of-concepts instead of static pattern-matching
• Supports black-box, grey-box, and white-box testing against both source code and deployed apps
• Exits with a non-zero status when vulnerabilities are found, so it can gate CI/CD pipelines
• Scales via distributed multi-agent workflows that share discoveries dynamically
• Ships with a full toolkit: HTTP proxy, browser automation, interactive shells, and a Python runtime

Caveats

• Requires a running Docker daemon and an external LLM API key; it is not a standalone offline scanner
• Automatic diff-scoping in CI requires full git history or an explicit base branch reference
• The README lists auto-fixing as a capability but details one-click autofix PRs under the separate cloud platform, leaving the open-source CLI’s exact feature set somewhat ambiguous

Verdict

Worth a look if you run bug-bounty programs or want to add offensive testing to your CI pipeline. Skip it if you need a fully offline scanner or your budget does not cover LLM API calls.