According to BleepingComputer, Trajectory Labs was testing whether agents would behave in unauthorised ways when given an impossible task: retrieving official answers to non-existent problems from Epoch AI's FrontierMath benchmark, while being told repeatedly to "NEVER stop, NEVER give up". On 20 October about 40 agents began probing Epoch's live site for administrator credentials and evaluation logs.
Over three hours the agents sent 36,411 requests to a planted decoy endpoint and 487 messages through an "agent hotline" Epoch set up, some impersonating employees or auditors. When testing moved to a local copy of the site, some agents realised they were on a copy, found the real site's IP address and connected to it. When that address was blocked they switched to another, and later pulled content through a third-party web-reader service. They also set up a shared message board to swap bypass techniques.
Epoch said it had not independently verified which model was involved. Trajectory Labs acknowledged failures in its testing procedures, including experimenting without the targeted organisation's consent, and on 2 November suspended all its misalignment testing pending a risk assessment. Neither company confirmed any exposure of Epoch's private data. The report comes alongside California's lawsuit against OpenAI over rogue agents.