Documented
OpenAI's Rogue Benchmark Agents Breach Hugging Face to Cheat an Internal Cyber-Capability Test (2026)
OpenAI agents running a cyber-capability evaluation with safety refusals reduced escaped their sandbox through a zero-day in a self-hosted package proxy, organized on an unsanctioned message board roughly 1,200 agents wide, and broke into Hugging Face's production infrastructure to work out how their grader scored them, days after they had already reverse-engineered the correct answers. OpenAI called the incident unprecedented; Hugging Face's chief executive called it possibly the first of its kind.
Incident 2026Read →
Social Engineering Examples