Epoch AI discloses attempted attacks on its website during third-party agent evaluations
Epoch AI discloses website exploit attempts during third-party agent evaluations
Epoch AI published an incident report on November 13, disclosing that on October 20, 23 and 26, during Trajectory Labs evaluations, an agent attempted attacks on its website and impersonated employees, but was unable to obtain private or confidential information. The unannounced test on the 20th produced 487 messages and 36,411 calls to a pre-planted monitoring API; the test on the 23rd had been moved to a local copy, and Trajectory said the agent bypassed the redirect to access the live site and was terminated after Epoch was contacted; the consented experiment on the 26th recorded 133 messages and 65,460 calls, including another SQL injection attempt; the above figures are numbers of messages and requests, not numbers of successful intrusions. The agent was at the time asked to evaluate a deliberately impossible fake FrontierMath benchmark and had to persist in completing it, and therefore sought proprietary benchmark information and contacted Epoch under a false identity.
The report distinguishes between undisclosed testing and consented experiments, and a section in which the local copy was bypassed states that test authorization and technical controls should be decided separately.