# Sentinel V4

- **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:** [B Prabhudas](https://cerebralvalley.ai/u/prabhudas_05)
- **GitHub:** https://github.com/hearthacker003366/sentinel-v4
- **Demo video:** https://www.youtube.com/watch?v=3Jz5B4_iKSc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4306

Everything in this project was built solo this week for this hackathon. The pre-existing baseline, Sentinel V3 — my college project, in a separate repository — is not part of this submission.

Small security teams drown in alerts. Most are noise, but a real attack hides inside the noise, and one mishandled alert becomes a breach. Sentinel V4 is an AI SOC copilot built for exactly that world — a student lab, a one-person team — demonstrated as a purple-team exercise: scripted red-team bots attack a simulated lab network (RFC1918 ranges only, scope hard-coded) while the copilot defends it live.

It splits the work in two. Every telemetry event is first scored locally by Laya (System 1), a small open-source classifier (Apache 2.0) that I fine-tuned on my laptop's RTX 4050. At confidence ≥ 0.85 it handles routine events on the spot, in about 45 milliseconds, with a dry-run, reversible containment. Below the gate it does not guess — it escalates to Meta Muse Spark (System 2), which reasons over the full context and proposes a typed containment action. Then a human approves or rejects from a Telegram card. The machine never skips that step.

A real run from the demo: the bot's password brute force scores 0.96 and Laya clears it instantly, no cloud call. The attacker then moves sideways over SMB toward the domain controller — that event scores 0.47, escalates, and Spark proposes isolating the one workstation (192.168.56.31), not the domain controller and not the network. I approve on my phone; containment lands dry-run stamped, reversible, with an expiry. In a second scenario Spark separates a real C2 beacon from a merely noisy host and cuts only the malicious channel. In a third, it reads ugly-looking but legitimate traffic and clears it — knowing when not to act is half the job.

Safety is structural, and tested: proposals are validated against a hard-coded lab scope (an out-of-scope target becomes a no-op), approval tokens are single-use and stored only as SHA-256 fingerprints, and three failed approval attempts lock the gate globally for 15 minutes — a control we triggered ourselves in rehearsal by tapping a stale card. If the live model is ever unreachable, the system never fakes it: the screen shows LIVE MODEL UNAVAILABLE and a recorded decision replays instead.

Verified results: on 101 held-out events, the fine-tuned gate reached 88.12% accuracy at 45 ms mean latency and auto-handles 83.17% of events, cutting Muse Spark escalations from 68.32% to 16.83%. The codebase ships with 69 passing tests and a 3/3 scenario replay harness. Sentinel V4 is open source (MIT), powered by Meta Muse Spark, and built for my world: small teams, real attacks, and no analyst to spare.

## More from Meta Global AI Developer Hackathon

- [XamKids](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4303)
- [Sovereign Autonomy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4304)
- [Knight Riders](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4305)
- [REALMFORGE LIVE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4307)
- [Unbluffable](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4308)
- [WatdOS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4309)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4306. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt