# NoKey — Business Protection

- **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:** [Vintimilla Palacios](https://cerebralvalley.ai/u/Bender)
- **GitHub:** https://github.com/bendervintimilla/bendrix-shield
- **Demo video:** https://youtu.be/FkB9mli3ZHU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/519

wAIrrior (submitted earlier as NoKey) is the AI warrior that watches every app, machine and camera a business runs, spots irregularities, and contains the attack on its own. Contain first, ask later: anything reversible it does alone in milliseconds; anything irreversible waits for the owner, who then gets one WhatsApp notice: "You were attacked. We already did X, Y, Z. Tap to undo."

This week: on 2 Nov attackers remotely forced supermarket freezers across Europe into defrost (Tom's Hardware, CBS, Al Jazeera); on 4 Nov a French bank suspended transfers for every client after fraudulent operations (Euronews); in October our own AI keys were abused for ~16 hours (USD 310).

Real cameras: Meta SAM 3.1 watches 5 real security cameras of our restaurant in Madrid. Every person is covered before any frame is shown. On 78 real frames labeled by hand: someone-vs-empty 78/78 (precision and recall 1.0), exact head count 59/78. Five security apps run on it: occupancy, after-hours intrusion, storeroom stock watch, unattended cash desk, and a shift report Muse Spark writes from redacted frames only.

Sentinel: one timeline for cold room, payments, POS and AI gateway. A combined attack wave is contained per app (defrost blocked, payment held, stolen key replaced) while legitimate work continues: in single-app attacks 300/300 legitimate writes stay alive vs 0/300 when everything is suspended.

Physics, not text: a defrost is a legitimate command, so before an agent acts a thermal twin predicts the product temperature. On 1,000 simulated scenarios the shield stopped 419/419 damaging actions (static rules 7/419; false blocks 23/581). A three-round red team published every bypass it found; each was closed.

Honest limits: cameras and Meta model calls are real; the attacks, the thermal twin and containment are simulated and labeled; no real actuator or bank is connected. Demo (no sign-in): https://bendrix-shield-panel-290240328973.us-central1.run.app/

## More from Meta Global AI Developer Hackathon

- [Iris](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/516)
- [SiliconMind](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/517)
- [ASTRO 6](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/518)
- [Aries Nyne](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/520)
- [Rajat(Self)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/521)
- [Pipeline Copilot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/522)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/519. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt