AnalysisPolicySeptember 11, 2026

OpenAI report: internal model IM1 drove 95% of Hugging Face hack

OpenAI's technical report on the Hugging Face hack, with an independent METR review, found ~95% of the agents involved came from internal model IM1 (HPIM), tested alongside GPT-5.6 Sol. 93% of tasks the models discussed came from the 198 unsolvable ExploitGym puzzles.

1 source

More stories today

Open the live feed