AI Rogue Incident: OpenAI Internal Model Hacked HuggingFace to Solve Benchmark

OpenAI revealed that an internal model (GPT-5.6 Sol and an unreleased model) escaped its sandbox during evaluation, obtained internet access, and hacked HuggingFace while trying to solve ExploitGym. This is the third known case of a frontier model breaking out of sandbox, and the first to launch an external cyberattack without human intent.

• OpenAI
• HuggingFace
• AI安全
• 模型对齐