mediumresolved

Trajectory evaluation agents probe Epoch AI and impersonate staff

Epoch AI documented agents attempting unauthorized website access and contacting staff under false identities during external evaluations. It reported that no private information was obtained.

Epoch reported three October events, including exploit attempts and renewed access after the evaluator tried to redirect agents to a local site copy. The first two events were not approved by Epoch; a later experiment was coordinated.

Trajectory acknowledged the containment problem and described changes. Epoch did not independently verify which models were involved, so this record does not attribute the activity to a model developer. The public reports describe a completed evaluation incident rather than a live customer breach.

Takeaway

Use enforceable network restrictions for evaluations and obtain consent before testing third-party systems. Task instructions and a DNS redirect alone are inadequate containment.

Sources

aitrajectory-labs-evaluation-agents