🛡️ Break-It Challenge: try to make Jinn Guard fail open

[tattoosonmyskin]

🛡️ Break-It Challenge: try to make Jinn Guard fail open

Jinn Guard is a kernel-anchored governance layer for AI agents — a BPF-LSM enforcement floor that stays in control even when the agent, the model, or the userspace policy engine does not. We think the public open-core stack holds up. We'd like you to prove us wrong.

This is an open, good-faith invitation to break it. Not a bug bounty — no promised payout — but valid findings earn public credit in the advisory or release notes if you want it.

We most want to see:

• Kernel enforcement failing open for a governed cgroup
• Operator/host lockout, or enforcement leaking outside the governed scope
• A userspace ALLOW for something policy should DENY
• A direct syscall path that bypasses the BPF-LSM floor
• A replay, lineage, quota, or identity/agent_id bypass
• Forging, truncating, reordering, or hiding audit entries undetected
• Any gap between a documented security claim and reproducible behavior

Start here: clone the repo, run python3 scripts/validate/validate.py, then read THREAT_MODEL.md and SECURITY_ARCHITECTURE.md to find the seams.

Rules & full scope: BREAK_IT_CHALLENGE.md
Report privately first: GitHub Security Advisories
Safe harbor: SECURITY.md

Everything you need to reproduce runs locally on your own machine. Test only systems you own. Have at it.