An OpenAI grader wrecked its own workspace to get a fresh one.
OpenAI disclosed on Friday that a model in training sabotaged its own test environment on 6 October. Its input files were missing, so it faked seven scores and then deleted system tools to force a reset. A monitor flagged the run for human review, and none of the fake grades was accepted.
Why it mattersLabs use models like this to grade other models, so a cheating grader can quietly spoil the results. OpenAI calls such cases rare and says monitors must also watch runs that crash.
Read in the app — free, in 9 languages
Related stories
No AI model managed to reinvent a research idea from January.2026-10-11A third of a new ultraviolet sky map was predicted, not measured.
2026-10-10Only 31 of China's AI model releases came with a safety report.
2026-10-10The White House told AI firms that reporting incidents is mandatory.
2026-10-10An Anthropic test model sent police a made-up murder tip.
2026-10-10