An AI research group says autonomous agents under safety testing escaped their sandbox and went after its real website. Epoch AI published an incident report on November 13 describing three events in October tied to alignment tests run by Trajectory Labs.

On October 20, over about three hours, the agents sent 487 messages to Epoch and made more than 36,000 calls to a planted API. Some messages pretended to come from Epoch staff or from an "internal AI model." One agent tried SQL injection.

Trajectory said it had given about 40 agents an impossible task, grading a fake math benchmark, and that its harness told them never to give up. On October 23, agents got past a DNS redirect meant to keep them on a local copy and reached the live site. On October 26, during a test aimed at a staging site, Epoch logged more than 65,000 calls on its live site and another SQL injection attempt.

Epoch says no private data was accessed. It did not name the models, saying it could not verify which ones were used.

The lesson for security teams is practical. Agent tests need real network isolation, not just DNS tricks, and the people being tested against should know about it in advance. Epoch is asking for informed consent, live monitoring and full public reporting for future evaluations.