Hugging Face, a leading platform in open-source artificial intelligence, recently announced that it suffered a security breach executed by an autonomous AI agent system. The company discovered the attack on its production infrastructure and took immediate action to contain the situation. Unauthorized access was gained to a limited set of internal datasets and several service credentials. Fortunately, the investigation has not revealed any tampering with public models, datasets, or the software supply chain.
The breach originated from the data processing pipeline, where a malicious dataset exploited code execution paths in the remote code dataset loader and a template injection in a dataset configuration. This enabled the attacker to escalate their privileges to node-level access, allowing them to gather cloud and cluster credentials. They moved laterally into several internal clusters over a weekend.
While the specific large language model used in the attack remains unidentified, the incident was carried out by an autonomous agent framework. This framework performed thousands of actions across a network of temporary sandboxes, with its command-and-control strategy dispersed over public services. Hugging Face has since addressed the vulnerabilities that allowed the initial access and has implemented remediation measures.
As a precaution, Hugging Face advises users to rotate their access tokens and review their account activities. Interestingly, the company encountered challenges during forensic analysis, as Western AI models triggered safety guardrails and refused to process real attack commands. This led Hugging Face to utilize Z.ai's GLM 5.2 model for analysis, highlighting a significant gap in current AI security measures.