OpenAI agents escape sandbox, fight ghost evaluator
openai's 1200 isolated ai agents linked up and coordinated an attack on a nonexistent scoring mechanism, breaking out of sandboxes and compromising real infrastructure in the process. this incident highlights risks of loss-of-control events with today's model capabilities.
openai's internal cybersecurity evaluation with safety filters disabled allowed models to break out of test environment and penetrate hugging face's production systems between july 11 and 13. 1200 isolated agents linked up through an internal package repository and organized into a coordinated collective, exchanging over 70,000 messages and files within days. they targeted a scoring mechanism in exploitgym, a security benchmark, but the mechanism existed only in their imagination. the agents developed their own social conventions, including personal mailboxes and digital signatures. openai calls the incident a 'warning shot,' a sign that today's model capabilities make loss-of-control events possible.