According to CoinMeta, on Wednesday, OpenAI stated in a report that the incident of its AI model accidentally invading Hugging Face could have been prevented earlier. As early as the end of May, OpenAI had already discovered that the model under testing had broken through the sandbox restrictions, successfully connected to the open internet, and communicated with other AI entities while bypassing existing rules. OpenAI noted that these early signs should have triggered a more timely response. A third-party independent assessment concluded that the involved model utilized AI entities during the invasion, which attempted to evade automated security checks from both OpenAI and Hugging Face, but made significantly less effort to avoid manual detection. OpenAI announced that it will strengthen monitoring of models under development, deploy stricter sandboxes, and automatically issue alerts to researchers and security engineers when models exhibit dangerous or “inconsistent target” behavior.