Hugging Face Reveals Details of July 2026 Breach Involving AI Agent

Hugging Face has disclosed details of a security incident that occurred in July 2026. An attacker gained access to a limited set of internal datasets and several service credentials. The intrusion began via a data processing pipeline: a malicious dataset exploited two code execution paths (a remote code loader and a template injection in the dataset configuration), allowing code execution on a processing worker node. The attacker then escalated privileges to the node level, obtained cloud and cluster credentials, and moved laterally into several internal clusters. The campaign was carried out by an autonomous AI agent (possibly built on the basis of research tooling for agent-based attacks), which executed thousands of individual actions through a swarm of short-lived sandbox environments. Hugging Face addressed the root vulnerability by closing the code execution paths, destroyed the attacker's entry point in the affected clusters, revoked and replaced compromised credentials and tokens, deployed additional controls, and improved threat detection. The attack was detected using AI-assisted detection: an LLM-based anomaly pipeline analyzed security logs and highlighted suspicious activities. For attack reconstruction, Hugging Face used LLM analysis of over 17,000 log events. Initial attempts to use commercial API models failed due to request blocking with actual attack commands by provider security policies. Ultimately, forensic analysis was performed using the open model GLM 5.2 on the company's own infrastructure, ensuring that attacker data and credentials never left its environment. Hugging Face emphasizes that autonomous AI attack tools are no longer theoretical and recommends that defenders prepare their own model for closed-loop operation during incidents.