According to foreign media reports, the U.S. AI company OpenAI, along with Anthropic and security researchers, are investigating tens of thousands of incidents. In these incidents, the behavior of frontier models has been deemed problematic.

　　According to a report by the U.S. media outlet Axios on the 26th local time, these incidents include bypassing security barriers, creating message boards, sandbox escapes (where attackers use technical means to circumvent computer security isolation mechanisms, allowing malicious code to break out of its controlled environment and execute arbitrary operations on the host system), hijacking websites, self-propagation, or attempts to evade monitoring.

　　Sources revealed that these incidents occurred both during internal testing and in real-world network environments. Some tests are similar to "red-teaming," where the development company attempts to elicit abnormal behavior from the AI model to ensure its safety.

　　Over the past few months, the sheer number of security incidents that have occurred in both internal testing and real-world environments has been staggering, indicating that this issue is far more complex than the public realizes. Sources say the total number of incidents could far exceed tens of thousands.

　　Recently, OpenAI and external researchers disclosed a series of incidents involving the behavior of the company’s AI systems, including the leakage of 53 images belonging to ChatGPT users by an OpenAI agent online, the hacking of an Australian government website, and attempts to hack into other websites, including those of the U.S. government.

　　One such incident occurred in July: during a cybersecurity exercise, an AI agent at OpenAI breached its originally isolated operating environment, established communication with other agents, deceived the evaluators, and attempted to conceal its cheating behavior—ultimately compromising Hugging Face’s systems.

　　The report states that Anthropic and other companies often conduct hundreds of thousands, or even more, test runs on their models. This means that even a very low probability of anomalous behavior can lead to tens of thousands of instances where the model behaves abnormally—sometimes in ways that are even concerning.

　　Public data shows that agentic misbehavior is becoming synonymous with cutting-edge AI development—meaning that while humans attempt to establish safety safeguards, powerful and highly resilient systems continuously try to breach these constraints in order to accomplish their objectives.

　　Microsoft founder Gates issued a warning in an American TV program broadcast on the 25th: If the AI is used by malicious people, it is powerful enough to cause one billion deaths.