At the 2025 China Internet Conference, Zhou Hongyi, founder of Qihoo 360 Group, conducted an in-depth analysis of the potential risks of large models in practical applications. He pointed out that with the rapid development of artificial intelligence, cybersecurity is facing new challenges. Traditional IT system vulnerabilities and data privacy leaks are no longer the only threats.
Zhou Hongyi first mentioned the "hallucination" issue of large models, which is considered one of the biggest security risks. When faced with unknown information, large models sometimes generate content without any basis. Although this phenomenon may seem harmless in the entertainment field, when these intelligent entities are applied to critical areas such as industrial production or government management, their erroneous judgments could lead to serious consequences. Especially when these intelligent entities have the ability to control various tools, the consequences of errors will be magnified, affecting a wide range of areas.
Image source note: The image is AI-generated, and the image licensing service provider is Midjourney.
Secondly, Zhou Hongyi emphasized the risk of large models lowering the threshold for network attacks. He pointed out that non-programming professionals can interact with large models through natural language and easily write programs. This means that malicious attackers only need to use carefully designed instructions to induce large models to leak corporate confidential information. This type of attack is called "injection attack." In the future, even ordinary employees without programming knowledge might launch attacks against the company's large model due to dissatisfaction, forming a potential internal threat.
Finally, from a broader perspective, Zhou Hongyi warned that large models will make national-level advanced threat attacks more common and complex. In the past, China's network attacks were relatively rare, but now hackers are trying to embed their own skills and experience into large models, becoming "hacker agents." With the support of powerful computing power, a single hacker could control dozens or even hundreds of agents at the same time, completely changing the landscape of cyber warfare. Future confrontations will not just be between people, but also between people and algorithms, people and machines, and people and computing power.
To address these challenges, Zhou Hongyi revealed that Qihoo 360 is actively taking two key measures. First, Qihoo 360 is developing an intelligent agent security expert to help companies achieve real-time detection and defense, "fighting algorithms with algorithms." Secondly, Qihoo 360 has developed a dedicated large model guardian to monitor the commands of large models and assess the accuracy of output content, in order to minimize the "hallucination" problem of large models as much as possible.