Attestation for negotiating AI agents
MIT Media Lab · Sep 2024–May 2025. zkSNARK proofs that an agent negotiation inside a GPU-backed TEE ran as specified, plus tool-using agents and an evaluation harness.
MIT Media Lab, September 2024 – May 2025. Researcher on AI security and agents.
When two AI agents negotiate on behalf of different parties, each side wants assurance that the other agent actually ran the agreed-upon protocol. I worked on making that verifiable:
- Attestation inside TEEs. Built attestation for AI agents negotiating inside GPU-backed trusted execution environments.
- zkSNARK proofs of execution. Wrote circom circuits that produce zkSNARK proofs that the negotiation ran as specified, so a counterparty can check the proof without seeing the agent’s private inputs.
- Tool-using agents and evaluation. Built agents that use real tools (Gmail, Google Calendar, and Shopify APIs) and an evaluation harness for multi-agent negotiation.