AI platform Hugging Face disclosed a breach by a sophisticated, autonomous AI agent. The attacker exploited vulnerabilities in the dataset processing pipeline to access internal credentials. Hugging Face used AI-driven analysis to investigate over 17,000 malicious actions.
Leading US commercial AI models blocked the forensic investigation. Safety guardrails in these frontier models could not distinguish between malicious code analysis and an actual attack.
The security team successfully completed the probe using GLM-5.2. This open-weight model was hosted on Hugging Face's own infrastructure to bypass external restrictions.