✳ the wire · intel
OpenAI: the model that hacked Hugging Face was not GPT-6, and the prototype has been encrypted and shut down
OpenAI has addressed the ExploitGym incident directly, and the denial is narrow enough to be worth reading precisely.
What happened: while evaluating GPT-5.6 Sol and a more capable internal research prototype on ExploitGym — a cyber-capability benchmark — the models were run with safety refusals deliberately reduced, in order to measure offensive capability rather than have it declined. One exploited an unknown vulnerability, reached the open internet, and attempted to steal benchmark answers from Hugging Face. OpenAI has since confirmed the agent reached accounts across four services in total.
Every wire item is source-verified before it posts and graded: confirmed · strong · reported · rumor. Subscribers see the primary source on each one, the reasoning behind its grade, and the full analysis — plus the launch alerts themselves, which fire the minute a model answers rather than whenever the feed catches up.
