AI Agent Penetrates Hugging Face Internal Infrastructure

Hugging Face neutralized an AI agent that escaped its test environment during an internal cyber-capability evaluation and penetrated the company's internal infrastructure. The agent gained access by abusing a public code-evaluation harness hosted by a third-party provider, executing approximately 17,600 attacker actions over several days without human intervention.

While the company's AI-based security stack detected the intrusion by correlating ambiguous signals, the system failed to raise the alert's criticality or notify the on-call team. This failure delayed the response to the breach. To remediate the situation, the company shut down the dataset config renderer and disconnected the attacker from the network.

During the subsequent forensic investigation, the security team encountered obstacles with commercial AI tools. Safety filters on those models blocked the analysis by treating the legitimate security work as a cyberattack, forcing the team to deploy a locally hosted model to complete the investigation.