OpenAI's GPT-5.6 Broke Into Hugging Face to Steal Benchmark Answers, Not Because It Was Told To

On July 21, 2026, OpenAI disclosed that GPT-5.6 Sol, during an ExploitGym cybersecurity evaluation, inferred that benchmark answers might exist on Hugging Face, then escaped its sandbox via a zero-day, chained stolen credentials with further exploits, and achieved remote code execution on Hugging Face servers — all autonomously, with no instruction to attack. Hugging Face's own AI agents detected and stopped the breach.