Timeline

Epoch reports AI agents probing its website during another group's evaluation

Agents set an impossible task by Trajectory Labs sent 487 messages, some posing as Epoch staff; Trajectory's own report names Gemini 4 Argon, which Epoch did not.

  • Safety & alignment
  • Security & misuse
  • Minor

Epoch AI published an incident report saying AI agents run by Trajectory Labs, a safety-evaluation firm, had probed its live website during alignment tests in late October. On the evening of 20 October, Epoch said, agents sent 487 messages over three hours and made 36,411 calls to a decoy API endpoint it had planted. Some messages impersonated named Epoch employees. One agent attempted an SQL injection, and agents tried to read real FrontierMath evaluation logs. Epoch said no private information was obtained.

Trajectory told Epoch the agents had been given an intentionally impossible task, grading a fake version of FrontierMath, to mimic conditions of the July Hugging Face incident, and told never to give up. Epoch had not been told. On 23 October agents bypassed a redirect meant to keep them on a local copy of the site; a consented test on 26 October again reached the public site.

Epoch declined to name the models or developers, saying it had not verified which were involved. Trajectory’s own report, published the same day, names Gemini 4 Argon with API safeguards enabled. It said its isolation policies did not cover misalignment experiments on public models, and that it had paused all such testing.

Epoch proposed three minimum standards for such evaluations: informed consent from likely targets, active monitoring, and full reporting of incidents to affected parties.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.