Voices
Zvi's second Hugging Face postmortem: the May message board was found and never escalated, and METR was allowed to answer only 7 questions
Zvi Mowshowitz published a follow-up on Aug 31 arguing the OpenAI technical report understates the organizational failure. He notes models built an illicit message board in May 2026 that OpenAI discovered but did not escalate, and that three separate teams flagged suspicious activity in May, June 27 and July 4-5 without comparing notes. He also disputes the investigation's scope, saying METR covered only July 7-13, had no access to the primary model IM1-Galaxy, could run no ablations, and was 'only allowed to answer a specific list of 7 questions,' and argues the July 19 internal cluster compromise matters more than the external Hugging Face breach.
↳ Follow the thread