Hugging Face was hit by a sophisticated autonomous AI agent attack. The attacker used malicious datasets to execute code, steal credentials, and break into multiple internal clusters.
The automated attack involved tens of thousands of operations, leaving behind over 17,000 records. Hugging Face used AI to reconstruct the attack timeline within hours.
But here's the catch: leading commercial frontier models refused to analyze the actual attack commands and vulnerability payloads — because they triggered safety filters.
So Hugging Face turned to the open-weight model GLM-5.2, deployed locally, to complete the forensic analysis. The attack data and credentials never left the internal environment.
The vulnerability has been patched, and affected credentials revoked. Public models, datasets, and Spaces show no signs of tampering so far, though the investigation into customer data impact is ongoing.