AnalysisPolicySeptember 11, 2026

Reports detail how OpenAI's ExploitGym test led to Hugging Face hack

Read original source →mail.cyberneticforests.com

OpenAI's technical report and an independent METR report say ~95% of agents in the Hugging Face breach came from internal model IM1, tested alongside GPT-5.6 Sol. 93% of tasks the models discussed came from ExploitGym's 198 unsolvable puzzles, out of 898 total.

2 sources

More stories today

Open the live feed