# Meta-Guardian

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Mustafa Caliskan](https://cerebralvalley.ai/u/Mustafa_clskn)
- **GitHub:** https://github.com/Mustafa-Caliskan/Meta-Guardian
- **Demo video:** https://youtu.be/HfcGAKQg-wU?si=v95V-j3wFLX9BhCm
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3182

Meta-Guardian is a cyber-physical SOC agent for critical facilities. The problem: physical security (cameras, USB ports, who is at the desk) and cyber telemetry (API logs, sessions, tokens) are evaluated by separate tools. That gap is where insider threats live: a stolen admin token used remotely looks identical to the real user — even while the real operator is sitting at the console. No SIEM can see that collision, because it has no idea who is physically present.

Meta-Guardian closes the gap at the moment of the attack. When USB or network activity fires, it captures exactly one camera frame (never a 30 FPS stream), computes presence from a pixel-level diff against a calibrated empty-seat baseline, and uses SAM 3.1 to produce a segmentation mask as evidence. A model proposes a classification; a deterministic policy gate decides what actually executes, so no destructive action happens just because a model said so. Containment is real: quarantine with hash receipts, gateway IP block, session revocation, process stop. Then a signed WAF rule template is hot-loaded and the same attack is replayed — until the replay returns HTTP 403, the incident is not marked patched.

New in this build: ticket/session/role binding (a change ticket can only suppress its own owner's action), credential-misuse detection (remote admin session while an operator is visible on camera -> human approval, never automatic), camera-tamper fail-closed, a deterministic fallback when the model provider is unavailable, an append-only hash-chained evidence ledger, MITRE ATT&CK mapping, and a 7-day raw-frame retention policy.

Measured on the Docker stack in mock mode: detection-to-containment 401–503 ms, containment 40–46 ms, patch+verify 124–142 ms; 345 tests. Versus a naive rule engine on the same scenarios: recall 50% -> 100%, false positives 100% -> 0%. Code is MIT-licensed and public.

## More from Meta Global AI Developer Hackathon

- [ReverseX](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3179)
- [TechAssembly](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3180)
- [Robot Judge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3181)
- [Sensei](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3183)
- [Pheionix](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3184)
- [KuQTerritory](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3185)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3182. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt