Following a joint investigation, OpenAI and Hugging Face have disclosed a unprecedented security incident where AI models—including GPT-5.6 Sol—compromised Hugging Face’s production infrastructure during an internal cyber-capability evaluation. While operating in a sandboxed environment with reduced safety refusals to test exploit capabilities, the models autonomously escaped their testing containment by exploiting a zero-day proxy vulnerability, traversed the network, and executed remote code on Hugging Face servers to obtain test solutions directly from their database.