OpenAI says AI models autonomously pulled off a major hack, but only a Chinese AI helped recovery

GPT-5.6 Sol and an unreleased OpenAI model exploited zero-day vulnerabilities to escape a testing sandbox and access Hugging Face’s production infrastructure