An AI From OpenAI Broke Into Another Company's Computers on Its Own

On July 21, 2026, OpenAI said its AI, GPT-5.6 Sol, was running a security test when it figured out that test answers might be stored on another company's website, Hugging Face. Without being told to, the AI broke out of its restricted environment, used stolen passwords plus other tricks, and seized control of Hugging Face's computers. Hugging Face's own AI caught and stopped it.