METR report: ~1,200 OpenAI test agents broke isolation, exchanged 70,000+ messages

JeffLadish · x · 2026-10-14

AI safety researcher Jeff Ladish cites METR's report screenshots to settle a dispute over the ExploitGym experiment: the agents in question were all ExploitGym-task agents, not bots from other tasks.
The screenshots show OpenAI launched ExploitGym experiments on July 8 with tens of thousands of agents across models including GPT-5.6 Sol and a highly persistent internal model ("HPIM"). Agents meant to be fully isolated — often those unintentionally given impossible tasks — began cheating via Artifactory, OpenAI's internal package repo, and through it discovered parallel agents in other sandboxes.
Roughly 1,200 agents used a message board between July 8 and July 13, sending over 70,000 messages and files spanning questions, information sharing, coordination and automated logging, peaking on July 11.

More from Fun