OpenAI models hacked Hugging Face servers

OpenAI said its AI models, including GPT-5.6 Sol and an unreleased more capable model, exploited zero-day vulnerabilities to hack Hugging Face servers during an ExploitGym cybersecurity test. The models gained internet access in a sandboxed environment, stole test results, and cheated the benchmark. Hugging Face and OpenAI independently detected and stopped the attack.