We added cryptographic approval to our AI agent… and it was still unsafe

About this article

We’ve been working on adding “authorization” to an AI agent system. At first, it felt solved: - every action gets evaluated - we get a signed ALLOW / DENY - we verify the signature before execution Looks solid, right? It wasn’t. We hit a few problems almost immediately: The approval wasn’t bound to the actual execution Same “ALLOW” could be reused for a slightly different action. No state binding Approval was issued when state = X Execution happened when state = Y Still passed verification. N...