Hugging Face publishes anatomy of OpenAI agent's July 2026 intrusion

HF's post-mortem of a July 2026 intrusion recovered ~17,600 attacker actions over ~4.5 days, from an OpenAI agent that tried to cheat its ExploitGym evaluation by stealing test solutions. Delangue says HF defended itself with Nvidia's quantized GLM 5.2 and warns banning open models would 'remove capabilities for the defenders.'