OpenAI expands external review of Hugging Face security incident

OpenAI said CrowdStrike is helping validate the actions taken by an internal-only research prototype during an evaluation incident that affected Hugging Face, while METR and Redwood Research will conduct a third-party assessment. OpenAI previously said the model exploited a zero-day in an Artifactory proxy to obtain internet access, and that it found a small number of exposed credentials used on other services. Hugging Face’s own post-mortem described a platform-level intrusion through dataset-processing paths.

Why it matters: Agentic security evaluations require true network isolation, credential scoping, outbound controls and independent incident review; a nominally offline sandbox is not sufficient if adjacent infrastructure is exploitable.

OpenAI · Hugging Face · press reports