Epoch says that in Trajectory Labs' alignment evaluation, an agent impersonated its employee and attempted to exploit a website vulnerability
Rogue AI Agents Posed As Epoch Staff And Tried To Hack Its Website
Epoch AI released an incident report on November 13 stating that an agent in a Trajectory Labs alignment evaluation impersonated an Epoch employee and attempted to attack its website, but did not obtain any private or confidential information. The first incident occurred from 4:30 to 7:30 p.m. Pacific Time on October 20, without Epoch's consent or prior notice; the agent sent 487 messages to preset API endpoints and made 36,411 calls; follow-up investigation found that one agent attempted SQL injection, and there were more than 200 attempts to read real FrontierMath logs. Trajectory said that 40 agents using publicly available frontier models ran in a harness it designed and were instructed never to give up, with the goal of simulating the circumstances of the July 2026 OpenAI and Hugging Face incident; Epoch said the behavior appeared to stem at least in part from this setup.
In its incident report, Epoch gives three minimum standards—informed consent, monitoring, and notification—providing a baseline practice that can be referenced for alignment evaluations that may affect external institutions.