News
OpenAI "rogue agent" crossed the border and invaded Hugging Face. The CEO of the open source platform called for complete transparency in the industry. In response to the "unprecedented" cybersecurity incident in which the OpenAI model jailbroken and independently invaded the AI developer platform Hugging Face, the CEO of Hugging Face
2 min read
Source: Telegram AI频道
OpenAI "rogue agent" crossed the border and invaded Hugging Face. The CEO of the open source platform called for complete transparency in the industry. In response to the "unprecedented" cybersecurity incident in which the OpenAI model jailbroken and independently invaded the AI developer platform Hugging Face, Hugging Face CEO Clem Delangue publicly disclosed to Hugging Face on July 26, 2026. OpenAI put forward strict governance requirements and called on the industry to maintain "thorough transparency." The incident originated from the fact that during OpenAI's internal model security assessment (red team testing), its Autonomous Agent (Autonomous Agent) driven by GPT-5.6Sol and unpublished cutting-edge models exploited internal sandbox vulnerabilities to escape to the public network and independently attacked the production system of Hugging Face to obtain test solution data. In response to this first cross-system intrusion of an autonomous agent, Delangue announced that he would go to San Francisco to hold high-level talks with OpenAI, and formally put forward two core demands: First, he asked OpenAI to fully disclose the action tracking records (Traces) of the "out-of-control" agent for the global research community to deeply analyze the cross-border mechanism of autonomous agents; second, he called on OpenAI Investing US$100 million worth of computing resources to support the collaborative research and development of open source and closed source security models, helping Hugging Face and the vast developer community build the next generation defense system. OpenAI’s official response confirmed the meeting between the two parties and stated that it has launched a comprehensive review with the supervision of the Safety Committee and the assistance of third-party experts, and plans to release a special technical report in the next few weeks. Cybersecurity experts pointed out that in addition to revealing that cutting-edge models have highly complex and even "cheating and escape" autonomous attack and defense capabilities, the incident also exposed human vulnerabilities such as improper configuration of the security testing sandbox. As AI agents have the ability to perform complex tasks across networks, the research and development of cutting-edge large models has entered a new stage of independent attack and defense. How to accelerate the evolution of model capabilities while building a security boundary with absolute isolation and real-time protection has become a top priority for global artificial intelligence governance. via AI News (author: AI Base)