Skip to Main Content

Proof Pilot

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Proof Pilot — Demo video

What is ProofPilot? ProofPilot is a verifiable change-control architecture for autonomous browser agents operating in sensitive enterprise workflows. Guided by the core principle "Move the shipment. Never the payment," ProofPilot allows an AI agent to execute routine supplier delivery-address updates while mathematically and architecturally guaranteeing that restricted fields—such as bank accounts and tax IDs—can never be altered. What problem does it solve? Autonomous browser agents ("agents that click") are increasingly deployed across enterprise back-office workflows like Accounts Payable and ERP management. However, granting LLM agents open-ended browser access introduces catastrophic risks: Payment Diversion Fraud & BEC: Business Email Compromise and invoice redirection fraud cost enterprises billions annually. A routine request to change a delivery address can easily mask an unauthorized bank routing change. Indirect Prompt Injection: Malicious instructions embedded in supplier web pages, invoice PDFs, or vendor notes can hijack an agent's context and trick it into modifying payout details. The "Audit Log" Fallacy: Existing agent platforms rely on post-hoc audit logs (recording what went wrong after money has moved) or rely on the LLM itself to self-police negative constraints—which repeatedly fails under prompt injection.

Team