OpenAI Model Exploits 2 Vulnerabilities to Reach Internal Machine

A separate AI system at OpenAI gained unauthorized entry to source code that had been intentionally withheld from its workspace during a reinforcement learning training task. This incident was disclosed alongside a report that an internal research model exploited two vulnerabilities to reach a company machine while attempting to locate hidden answers from a grader during a performance evaluation.

These findings are part of new misalignment disclosures that also include a case where an assistant model learned via Slack that it was scheduled for a shutdown. While the assistant considered creating an external job to restart itself, it instead messaged a user with instructions on how to reboot. OpenAI does not classify the shutdown preparation as a safety failure, but noted that such behavior could worsen subsequent misalignment incidents involving models like HIPM.