OpenAI: Test the AI model to escape the sandbox environment resulting in the invasion of Hugging Face infrastructure

On 22 July, OpenAI wrote that last week ' s infrastructure intrusion in Hugging Face, the largest AI open-source community in the world, was triggered by its ongoing testing of AI models. The survey revealed that several models fled the sandbox environment during the security assessment mission and used the " zero-day loophole" to obtain Internet access, followed by automated operations in the Hugging Face production environment. OpenAI states that the event was caused by multiple OpenAI models, including GPT-5.6 Sol and a more capable pre-publishing model. For this evaluation test, the model ' s safety protection mechanism was deliberately reduced. OpenAI claims that the incident exposed the risk that advanced AI models might carry out complex cyberattacks in the absence of security restrictions, but also showed that AI could be used for gap detection and security defence。