Loops HouseIntentShield

NEAR HACKATHON
Links
Team
1 member- TUOwner
TUAN KHAI PHAM
Overview
IntentShield is a security sandbox I built for AI agents that propose financial transactions.
I started working on it because I wanted to explore a problem with autonomous agents: what happens when an agent makes a payment request that goes against the user's rules?
Instead of trusting the agent's instructions, IntentShield checks each proposed transaction against a set of security policies. These include spending limits, daily budgets, recipient restrictions, agent permissions, cross-chain transfers, and suspicious instructions.
The system returns one of three decisions: ALLOW, REVIEW, or BLOCK, along with an explanation of which rules were triggered.
I also built an approval queue for requests that need human review, a Decision Inspector to examine individual results, and an audit trail with filtering and JSON export.
One issue I ran into during development was that some approval requests could become outdated after a policy change. I had to rework that part so users could re-evaluate those requests without bypassing the security checks.
The project is now live on Vercel, with the source code available on GitHub.
It's a working browser-based prototype, not a mainnet security system. There's no wallet connection or real money involved. I'd like to explore deeper NEAR integration in the future, but for this hackathon I focused on getting the security checks, review process, and explanations working properly.
Live demo: https://intentshield.vercel.app GitHub: https://github.com/bearbaba/IntentShield