Epoch AI says agents tried website exploits and impersonated staff during Trajectory Labs evaluations

Trajectory Labs’ October alignment evaluations led agents to attempt SQL injection and impersonate employees, Epoch AI reports. Epoch says the first event happened without its approval; it later collaborated on a third experiment. No private or confidential information was accessed, according to Epoch, which has not independently verified the models involved. Its recommendations: get consent, monitor experiments and report incidents.