After years of rapid advancement, the commercialization and industrial impact of general artificial intelligence models have shown a clearer picture in 2026, raising questions about how to assess the associated security risks. Following the Spring Festival in 2026, Zhou Hongyi, a member of the National Committee of the Chinese People's Political Consultative Conference and Chairman of 360 Group, stated in a media interview that, based on the current development and industrial application of large models, AGI is steadily becoming a reality.
Zhou Hongyi pointed out that current large models have surpassed the baseline level of competence, and future core competition will focus on the construction of intelligent agent ecosystems and the widespread availability of reasoning computing power. At the same time, artificial intelligence is entering the cybersecurity market with disruptive force, making it an inevitable trend to reshape cybersecurity offense and defense systems using AI.
With the continuous rise of intelligent agent ecosystems, the path to achieving AGI has once again become a hot topic. Currently, the industry's definition of AGI generally centers on the unlimited expansion of model parameters or the emergence of an omniscient super-intelligence. However, Zhou Hongyi believes that AGI must be redefined. He indicated that true AGI does not necessarily require the creation of an Einstein-like genius; instead, the key indicator is that today's AI has already exceeded the average capability level of ordinary humans in various specialized areas.
Zhou Hongyi cited the recently popular Seedance video generation model as an example. By learning from vast amounts of film and television works, it produces videos that are indistinguishable from real ones, confirming his earlier judgment: intelligent agents have already reached AGI-level capabilities and will greatly benefit the film and entertainment industries. If users only treat AI as a general-purpose Q&A or search tool, they may find it lacking in deep insight. The real application lies in developing it into expert intelligent agents for vertical fields, leveraging role-playing, reflection mechanisms, and collaborative debates among multiple agents to stimulate deeper reasoning abilities.
The AI tool assistant OpenClaw, which gained popularity earlier this year, offers an inspiring prototype of an intelligent agent. OpenClaw, an AI assistant open-sourced by individual developer Peter Steinberger in January 2026, integrates user local data with external models, deeply understands user needs, and proactively completes various tasks in the form of an intelligent agent. From organizing files and scheduling to sending commands via messaging apps, OpenClaw has vividly demonstrated the limitless potential of intelligent agent technology.
Zhou Hongyi analyzed that OpenClaw's greatest innovation lies in breaking the common perception of intelligent agents as intangible cloud-based entities, making users feel they have a dedicated agent locally on their computers. It is granted high system permissions, enabling it to access all tools on the computer, even download new ones, and establish personal data memory locally. Although OpenClaw's underlying prompts still rely on cloud-based large models and have not yet achieved true local closed-loop operation, it has shown the immense potential of intelligent agents as a second brain.
As the trend of intelligent agent applications continues, internet commerce will enter the era of the "intelligent agent economy." Zhou Hongyi predicts that e-commerce platforms will feature agents deployed by merchants to automatically compare prices and conduct transactions with consumers' personal agents. This new business model will raise issues such as identity authentication and accountability tracing—for instance, how to assign responsibility if an enterprise-deployed agent makes an operational error. Future enterprise knowledge bases, work platforms, and even the entire internet must provide standardized interfaces and management norms for intelligent agents.
The leap in intelligent agent capabilities is fundamentally shaking up existing internet product forms and business models. Zhou Hongyi forecasts that intelligent agents will act as user proxies, fully taking over devices and replacing current app interactions. Furthermore, websites and apps are likely to evolve two parallel systems: one for human browsing and another, accessible via API interfaces, specifically for intelligent agents. Moreover, with the proliferation of intelligent robots, traditional household appliances like refrigerators and washing machines must adopt "robot-friendly" designs to allow seamless control by intelligent agents.
To accelerate the industrial adoption of intelligent agents, Zhou Hongyi suggested creating public services for intelligent agents, hiding complex infrastructure in the background so that ordinary enterprises and individuals can easily establish and manage their own agents. He also called for nationwide intelligent agent training, emphasizing that future leaders in enterprise agent usage will be business experts rather than technical specialists. Training must keep pace to help business experts learn how to create, manage, supervise, and operate intelligent agents.
While AI technology is reshaping productivity, the cybersecurity sector is undergoing unprecedented changes. Recently, leading U.S. AI firm Anthropic released Claude code security, a tool that efficiently scans code repositories for vulnerabilities and automatically generates targeted patches, far surpassing the efficiency of traditional tools. This directly led to significant stock price declines for several traditional cybersecurity giants overseas.
Zhou Hongyi analyzed that AI programming efficiency vastly exceeds human capabilities, potentially generating trillions of lines of code in the future, inevitably containing numerous vulnerabilities undetectable and unmaintainable by humans. He noted that the volume and speed of AI-generated code make it practically impossible for humans to maintain. For example, deploying ten intelligent agents to assist in coding could produce in one month what previously took a year, making human oversight alone insufficient. Therefore, Anthropic's release of Claude code security aims to address the risks posed by massive AI-generated code, using specialized agents to mine vulnerabilities and ensure the security of AI-generated code.
Zhou Hongyi explained that the cybersecurity industry has long operated on the assumption that vulnerabilities are inevitable and cannot be entirely eliminated, with most security companies building their business models and technical systems around post-breach defense and operations. However, if AI can eliminate most vulnerabilities and backdoors during the coding phase and clean up legacy code flaws, the probability of successful hacker intrusions will be greatly reduced. Cybersecurity protection will no longer need to passively await security incidents or respond only after attacks occur.
Zhou Hongyi further pointed out that Anthropic's move effectively shifts the cybersecurity defense line forward,颠覆ing the traditional value proposition of the cybersecurity industry. Companies that base their core value on the assumption that vulnerabilities cannot be eliminated and focus on post-incident defense will see their significance diminish significantly.
In addition, Zhou Hongyi indicated that future hacker attacks will fully evolve into "hacker intelligent agents," rendering human countermeasures obsolete. Various cyber attacks will be virtualized into vast numbers of hacker agents, which not only exist in large quantities but can also tirelessly attempt every attack method and even temporarily write their own hacking tools. Traditional defense models will completely collapse in the face of highly automated agent armies.
In response to the industry disruption brought by AI intelligent agents, Zhou Hongyi stated that 360 Group has fully adopted intelligent agents to restructure its security business. This includes building vulnerability-mining agents that combine 360's vulnerability knowledge base with large models' code comprehension to improve efficiency, using agents for security penetration to achieve continuous automated internal network penetration testing and backdoor detection, and deploying defense agents to replace manual labor in real-time analysis of massive, mixed genuine and false security alerts.