# Zyphers

- **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:** [ABISHEAK S](https://cerebralvalley.ai/u/_abisheak_s)
- **GitHub:** https://github.com/S-ABISHEAK/intent_lock
- **Demo video:** https://www.youtube.com/watch?v=MV2n0Z0dA2Y
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2415

AI shopping agents can scan product pages and product reviews and emails can include hidden instructions. A bad review can move an agent to purchase something you don't want, order an extra with your purchase, send your purchases to someone else's address, or give your cart to an attacker via email. And your instructions cannot be differentiated from the attacker's by the agent.

IntentLock is positioned between the agent and the store. You make your statement in simple English, say, "a laptop under ₹60,000, ship to home". IntentLock makes that a requirement that you affirm. From then on, all the agent can do is make suggestions. IntentLock contains the merchant's true price and reviews each and every offer in relation to your mandate and a collection of hard-coded rules. It also logs the origin of each value: If the SKU was provided by a review then IntentLock is aware of it; if it was provided by you, then IntentLock is also aware of it. Safe actions pass through and risky is immediately set to an approved card that is visible in your screen, and anything that violates a hard rule is blocked. Approvals are cryptographically signed, limited to a once-only use and have a timeout of 120 seconds. Each step is recorded in a tamper proof receipt log.

Muse model is an helper inside Meta, that drafts the mandate, reads the checkout screenshot and flags the injected text. Only deterministic rules can tell "yes". During testing, 0 out of 46 attacks succeeded, all 34 normal purchases went through, and no additional prompts were required for normal purchases. If it's not in the line of sight, all 46 attacks get through without protection.

The agent comes with the suggestion, IntentLock decides.

## More from Meta Global AI Developer Hackathon

- [Vigilens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2412)
- [TeamHsain](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2413)
- [Nikola Reljin](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2414)
- [Solo Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2416)
- [Quoc Vu Junior](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2417)
- [DigitUP](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2418)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2415. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt