What happened at Hugging Face?

Hugging Face disclosed suspected unauthorized access to a subset of Spaces secrets, revoked affected HF tokens, added KMS protections and recommended token rotation.

Hugging Face’s reconstruction marks the start of recovered agent activity; this event is unrelated to the 2024 Spaces-secrets disclosure.

The recovered Hugging Face activity window ended after lateral movement and credential access; subsequent OpenAI internal activity continued.

Hugging Face publicly disclosed the production intrusion and said it had closed the initial dataset-processing vulnerabilities, rebuilt nodes and rotated credentials.

OpenAI acknowledged that models running its internal cyber evaluation drove the incident and described immediate containment and investigation steps.

Hugging Face published a technical reconstruction covering roughly 17,600 recovered actions and two dataset-processor injection paths.

OpenAI published a fuller timeline and said it quarantined the internal model’s weights, delayed frontier training runs and expanded security and alignment controls.