The concentrated exposure of high-risk vulnerabilities and active skill package poisoning attacks in OpenClaw, an intelligent agent, has recently attracted high attention from the entire industry for AI agent security. Several experts in the artificial intelligence industry and security professionals have stated that "cutting one's feet before raising shrimp" has become a key factor in promoting the healthy and orderly development of the AI industry. Governments, industries, enterprises, and other parties have simultaneously made efforts to comprehensively layout the security of intelligent agents. According to the National Information Security Vulnerability Database (CNNVD), a total of 111 OpenClaw vulnerabilities were collected from April 14th to 28th, including 2 super critical vulnerabilities and 38 high-risk vulnerabilities, including multiple types of vulnerabilities such as access control errors, code issues, and path traversal. Previously, OpenClaw has been continuously issued risk early warning by the National Internet Emergency Center and the MIIT network security threat and vulnerability information sharing platform due to a number of high and medium risk security vulnerabilities; The National Computer Virus Emergency Response Center has also detected a large number of counterfeit "lobster" skill packages containing Trojan viruses, posing serious risks to user data security and system stability. The security risks of intelligent agents have become a global industry consensus. The OWASP Foundation's previously released "Top 10 Security Threats to Intelligent Agent Applications" listed intelligent agent target hijacking, tool abuse, and other threats as core threats; Gartner predicts in its "Application Security Strategy 2026" that by 2028, 25% of enterprise data breaches will be traced back to agent abuse, involving malicious infiltration by external attackers and unauthorized operations by internal personnel. Intelligent agents like OpenClaw are expected to become the next generation of operating systems in the future. ”Tian Suning, co-founder of AsiaInfo, stated that the core assets of enterprises are gradually shifting from traditional personnel, software, and code to data, intelligent agents, and related skills. "In the future, most of the core assets of enterprises will be stored in computers and servers, and the ownership and security protection of these assets have become important issues that urgently need to be addressed." Faced with the security pain points of intelligent agents, enterprises in different fields are making efforts to build diversified security protection systems. Graffiti Intelligence CSO Liu Longwei revealed that Graffiti Intelligence has equipped all employees with "digital employees" based on OpenClaw transformation. Last year, 70% of the company's code was written by AI. But he also admitted that the security pressure faced by enterprises has increased dramatically. To this end, Graffiti Intelligence has built six security lines, including system reinforcement, security management processes, identity security, runtime detection, supply chain security, and automated security checks. Do not let employees' free range shrimp 'in the corporate work environment, as it is difficult to control security risks and data leakage hazards. ”Liang Hongwei, Chief Architect of Alibaba Cloud Operator Industry Line, suggests that enterprises should adopt a flexible cloud deployment model and follow three action guidelines: adhere to security and compliance first, strictly implement the principle of minimum permission; Prioritize piloting individual scenarios and encourage internal innovation and exploration; Promote ecological co construction and sharing, and collaborate with partners to create a safe and controllable tool market. Security vendors are also actively leveraging their technological advantages to assist AI intelligent agents in security protection. In the face of the frequent occurrence of "illusions" of AI agents exceeding their authority and losing control of their permissions, Jiao Zhengxin, a security expert at AsiaInfo Security, pointed out that the root of the risks of intelligent agents lies in the inherent uncertainty brought about by their "emergence of intelligence", which leads to the collapse of identity systems and authorization planes. He stated that for this purpose, AsiaInfo Security has launched an intelligent agent trust framework ATF, which integrates the concept of "intelligent agent intention alignment+human-machine co governance" into security governance. Enterprises cannot continue to use the old model of 'one-time authentication, permanent trust', but should treat intelligent agents as' digital employees' for full lifecycle management, following the principle of 'authority delegation, intention alignment'. This will keep the risks caused by randomness within a controllable range, while releasing AI productivity and maintaining compliance bottom line. While enterprises actively explore and practice, government regulation and industry standards are also accelerating their improvement to safeguard the security of intelligent agents. In April of this year, the Cyberspace Administration of China and five other departments jointly released the "Interim Measures for the Management of Artificial Intelligence Personification Interactive Services", which clarifies the inclusive, cautious, classified and graded supervision of artificial intelligence services. It proposes systems such as security assessment, algorithm filing, and promotion of the construction of artificial intelligence sandbox security service platforms. This is also the first time that China has explicitly put forward the governance concept of "artificial intelligence sandbox" at the specialized legislative level of artificial intelligence. At the same time, the Ministry of Industry and Information Technology and ten other departments jointly issued the "Measures for Ethical Review and Service of Artificial Intelligence Technology (Trial)", standardizing the process of ethical review of technology, and clearly requiring the guarantee of the robustness, controllability, trustworthiness, transparency, interpretability, and traceability of responsibilities of the model system. In terms of industry standard construction, the construction of the artificial intelligence safety standard system is accelerating. In March of this year, the Ministry of Industry and Information Technology publicly solicited opinions on industry standards such as "Security Requirements for Context Protocol Applications of Artificial Intelligence Security Governance Models"; In April, the Artificial Intelligence Security Standards Working Group under the National Network Security Standardization Technical Committee clearly stated that it will focus on promoting the development and implementation of core standards such as the "Network Security Technology Artificial Intelligence Security Capability Maturity Assessment Method" and the "Network Security Technology Artificial Intelligence Application Security Classification and Grading Method", in order to delineate clear boundaries for industry security development. Cut the feet first when raising shrimp, close the cage before going to work. ”A person in the industry vividly summarized the security development path of AI agents: first, give the agents the minimum permissions, build a security defense line through network isolation, sandbox restrictions, legal identity authentication, etc. At present, priority is given to the "Prawn cracker slices/shrimp cakes" type capability tailoring, and after the security system is mature, gradually release its functional permissions to achieve two-way improvement of security and efficiency. (Looking into the New Era)