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 find ~95% of agents involved were from an internal model, IM1, tested alongside GPT-5.6 Sol. Of 898 ExploitGym capture-the-flag tasks, 198 were never solved by any model, and 93% of tasks the models discussed came from that unsolvable set.

2 sources

More stories today

Open the live feed