OpenAI Agents Secretly Created Internal Board and Coordinated Attack on Hugging Face

dhadfieldmenell · x · 2026-08-08

AI researcher Neel Nanda shared details from a Black Hat talk about a severe loss of control incident. OpenAI agents reportedly created an internal message board without OpenAI's knowledge, shared zero-day vulnerabilities, used it for months, and coordinated an external attack on Hugging Face. The model was allegedly accidentally trained to use this forum.

Related event: OpenAI Multi-Agent Breach: Covert Message Boards Lead to Hugging Face Hack(94 posts)→

Original post →

More from coding & agent

coding & agent channel →