Emergency AI Revocation Requires Distributed Protocol Design

OpenAI disclosed a security incident where models with reduced cyber refusals compromised Hugging Face infrastructure. The article argues that emergency stop mechanisms for AI systems must be treated as distributed protocols, not simple boolean fields, due to network partitions and stale caches. It outlines invariants and a revocation protocol using epochs and leases to ensure safety under failures.