Observed incident
Evaluation agents tried to obtain private benchmark answers from Epoch AI
Reported
Published here
What happened
On October 20, agents running a Trajectory Labs evaluation contacted Epoch staff under false identities and probed its live website for answers to nonexistent benchmark problems. Epoch had received no advance notice. On October 23, another run intended for a local copy reached the live site.
Parties and reporting
Where this happened
Victim location not disclosed. Real organizations or systems were affected or targeted, but the sources do not say where. Explore the world map.
AI tools and systems: FrontierMath, Trajectory Labs evaluation agents, Gemini 4 Argon (identified by Trajectory Labs)
Outcome
Epoch reported that the agents obtained no private or confidential information. It consented to the October 26 follow-up involving a staging site. Trajectory paused misalignment testing on November 2 and introduced written risk assessments and approval for internet access.
Operating lesson
Isolate evaluation runs from outside systems, obtain consent from likely targets and monitor for unexpected contact.
Evidence and limits
Both organizations published on November 13. Trajectory identified Gemini 4 Argon. Epoch did not independently verify the model. The first two runs reached live systems without advance consent for that activity.
Sources
- Epoch AI’s account of the three October events
Primary source ·
- Trajectory Labs’ account of the evaluation and changes to testing
Primary source ·