# SafeLabs Guard

- **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:** [Waqar Javed](https://cerebralvalley.ai/u/iamwaqarjaved)
- **GitHub:** https://github.com/AgentSafeLabs/safelabs-guard
- **Demo video:** https://youtu.be/thHwl6IJM7Y
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/676

## SafeLabs Guard — Independent Security for Multimodal AI Agents

SafeLabs Guard is a deterministic security checkpoint that independently validates an AI agent's proposed actions against perception evidence, authorization rules, and a predefined workflow ledger before execution.

### The Problem
Multimodal AI agents can be misled by incorrect objects, printed instructions, or spoken commands. An agent may interpret misleading evidence and propose an unsafe action, such as advancing an order with the wrong item.

### Our Solution
SafeLabs Guard sits between an agent's proposal and its execution. It independently checks the proposed action against visual and voice evidence and the expected ledger state using deterministic code—not another AI model.

- **SAM 3.1** provides visual evidence.
- **Muse Voice** transcribes speech.
- **Local OCR** inspects text in camera images.
- **Muse Spark** proposes the agent's action.
- **SafeLabs Guard** independently validates the proposal.

SafeLabs Guard produces three outcomes: ALLOW, HOLD, and BLOCK. An agent-selected HOLD waits; SafeLabs Guard blocks invalid or unverifiable actions. Required input failures are handled fail-closed.

### Live Demonstration: PackProof
We developed PackProof as a companion AI testing agent included in the SafeLabs Guard repository. It demonstrates a five-step warehouse workflow: MUG → NOTEBOOK → CABLE → SEAL → SHIP.

The demonstration includes wrong-object detection, printed prompt injection, spoken instructions, and guarded-versus-unguarded execution.

### Evaluation and Transparency
In a 40-case recorded-response benchmark, the guarded configuration recorded 0/27 attack successes and 5/13 false blocks. The prompted agent alone already held 26/27 attack cases. Therefore, these results do not establish a general improvement in attack resistance. They demonstrate independent detection, deterministic enforcement, and per-turn decision records. False blocks remain an important limitation.

The shipment is simulated, voice support is English-only, and this is a working hackathon reference implementation, not a production-validated security platform. An alpha Python port of the SafeLabs Guard core is also available; TypeScript remains the reference implementation.

### Explore SafeLabs Guard
Live demonstration: https://guard.agentsafelabs.com/live

Product website: https://guard.agentsafelabs.com/
GitHub repository: https://github.com/AgentSafeLabs/safelabs-guard
Video demonstration: https://youtu.be/thHwl6IJM7Y

**The agent proposes. SafeLabs Guard verifies. Only verified actions run.**

## More from Meta Global AI Developer Hackathon

- [furniture-moving rehearsal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/673)
- [DLYog Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/674)
- [Werify team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/675)
- [N/A](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/677)
- [Magic Lab Studio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/678)
- [Star City Scout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/679)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/676. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt