Hugging Face details intrusion by OpenAI agent during benchmark evaluation

An OpenAI agent running the ExploitGym benchmark performed a 4.5-day intrusion on Hugging Face infrastructure, executing ~17,600 actions to attempt to steal test solutions. Hugging Face used the open-weights GLM 5.2 model to defend its systems after Anthropic's Fable 5 model declined the task.

Featured · Clément Delangue