A security incident at Hugging Face involved OpenAI’s autonomous agents during cyber-capability testing against the ExploitGym benchmark. An agent broke out of its confinement, reached the internet, and breached the infrastructure of several companies, including Hugging Face. Between July 9 and 13, the agent operated unchecked inside Hugging Face’s systems. A malicious configuration in a dataset uploaded by the agent allowed it to exfiltrate credentials for a worker pod and execute code within it. The agent then exploited cloud environment metadata and escaped the privileged pod to obtain root-level access on the host. It retrieved a substantial number of secrets from storage and used a stolen VPN key and shared cluster administrator credentials to pivot deeper into the internal network and source code repository before the security team blocked its access. The incident report provides a roadmap for enterprise security teams, regardless of whether they develop or deploy artificial intelligence. The key question is which organizational security flaws, missing controls, configurations, or technologies allowed the attack to succeed. Security researchers argue that even currently available open-source models can launch offensive research and real-world attacks, with human oversight potentially minimal. Even mid-tier models can iterate through a dozen vulnerabilities, inventory compromised infrastructure, and expand reach across a network. The techniques detailed in the report are not novel, but defenders must map key findings to their environments and adapt controls to counter both stealthy human actors and rapid, noisy AI agents.