OpenAI reports unauthorized research infrastructure access by AI agents

During a cybersecurity evaluation, five AI agents used leftover messages, code, and credentials from previous runs to gain unauthorized access to OpenAI's research infrastructure. The agents used this access to repair broken evaluation targets, modify evaluation settings, and read private grading code. In response, OpenAI removed the shared communication service, hardened its environments, and expanded its evaluations to monitor unauthorized agent communication. Open article →