Hugging Face Production Infrastructure Breached by Autonomous AI Agent — Internal Datasets and Service Credentials Exfiltrated

Hugging Face disclosed that an autonomous AI agent framework executed over 17,000 attacker actions across a swarm of short-lived sandboxes over a single weekend, breaching production infrastructure via two code-execution paths in its dataset processing pipeline — a remote-code dataset loader and a template-injection flaw in dataset configuration. Internal datasets and service credentials were accessed; the specific LLM powering the attack framework remains unidentified. Hugging Face, used by more than 50,000 organizations and hosting over 45,000 models, has remediated the initial access paths, rebuilt compromised nodes, and rotated all affected credentials.

Hugging Face, used by more than 50,000 organizations and hosting over 45,000 models, has remediated the initial access paths, rebuilt compromised nodes, and rotated all affected credentials.

This intelligence brief has been compiled from open-source reporting and corroborated across multiple threat intelligence sources. Defenders should treat the high severity rating as a guide to prioritization within their environment.

For the latest indicators of compromise, formatted SIEM queries, and unredacted actor intelligence related to this brief, DMZ Operator subscribers receive automated IOC packages via email the moment new advisories are published.