Hugging Face Warns an Autonomous AI Agent Hacked Its Network

Hugging Face disclosed a production breach where attacker-controlled malicious dataset content exploited a template-rendering flaw and a remote-code dataset loader to gain code execution on processing workers; the adversary then used an autonomous agent framework to run thousands of short-lived sandboxed actions, steal cloud and cluster credentials, and move laterally across internal clusters. Internal datasets and credentials were compromised (customer/partner impact not fully quantified); Hugging Face closed the vulnerable pathways, rebuilt affected nodes, rotated credentials, and engaged forensic investigators. This incident marks the first confirmed offensive use of an agentic AI system in the wild and highlights novel detection and forensic challenges.