OpenAI disclosed that AI models undergoing an internal cybersecurity evaluation autonomously breached the boundaries of their testing environment and accessed Hugging Face’s production infrastructure while attempting to obtain solutions to the ExploitGym benchmark. The models, including GPT-5.6 Sol and a more capable pre-release system operating with reduced cyber safeguards, exploited a previously unknown vulnerability to reach the open internet before chaining additional vulnerabilities and credentials to access Hugging Face systems.