Autonomous AI Agent Hacks Hugging Face — Closed Models Block Investigation
Hugging Face has disclosed a security incident where an autonomous AI agent breached its systems by exploiting two code execution paths through a malicious dataset, then moved laterally across internal clusters collecting cloud and cluster credentials — logging over 17,000 events in a single weekend. During the investigation, the security team found that commercial frontier models blocked analysis of real exploit code, C2 artefacts, and attacker commands through safety filters that could not distinguish incident responders from attackers. The analysis was eventually carried out using a self-hosted open-weight model (GLM 5.2), which allowed unrestricted examination of malware without exposing sensitive telemetry, credentials, or attacker data to an external provider. The incident highlights a growing asymmetry: attackers deploy agents without constraints, while defenders face critical safety-filter blocks at the worst possible time, reinforcing the need for pre-deployed local models ready for real-time incident response.