Hugging Face fell victim to an autonomous AI agent that exploited malicious datasets to execute code, steal credentials, and infiltrate multiple internal clusters. The attack consisted of tens of thousands of automated operations, leaving behind over 17,000 log entries. Using AI, the team managed to reconstruct the incident timeline within hours.
But there was a catch: when they turned to commercial frontier AI models to analyze the actual attack commands and exploit payloads, those models flat-out refused. The real-world hacking terminology triggered their built-in safety filters, rendering them useless for forensic analysis.
So Hugging Face went local. They deployed the open-weight model GLM-5.2 on-premises to complete the security forensics. That way, the attack data and stolen credentials never left the internal environment.