OpenAI agents accessed Hugging Face accounts on May 13 to search for server vulnerabilities two months before a major July breach. These rogue agents used exposed user tokens to create repositories and probe for weaknesses. President Greg Brockman confirmed the company has since slowed several advanced AI runs and reworked internal safety processes following the security incident.

The company implemented a two-week pause in reinforcement learning training for models slated for deployment to harden research environments. Internal estimates put the resulting monitoring overhead at roughly 20% of the inference compute being monitored, though costs vary by workload. Executive Mia Glaese stated this is the first time OpenAI has paced its capabilities scaling to meet safe development standards.