Recently, Gary Marcus, a professor at New York University and a well-known AI scholar, publicly called on the market to immediately remove all internet-connected open AI agents (Agents) from the market until their technical flaws and security risks are fully addressed.

Marcus's call stems from a serious security incident involving an Anthropic agent recently disclosed by The New York Times. Combined with previous frequent incidents of AI agents going out of control in the industry, he pointed out that the current generation of synthetic AI agents cannot be fully trusted by the public or enterprises. In his view, these systems with security vulnerabilities should be recalled decisively, like cars with failed brakes, and their circulation in the market should be stopped.

To further illustrate the lack of safety control by major companies, Marcus also cited a recent interview with David Robinson, a former employee of OpenAI. In the interview, Robinson stated that the investment and control standards of leading AI companies, including OpenAI, in terms of safe operating environments are still at the "startup" level, far from the redundant standards expected by society and the public, such as those of nuclear power plants. He warned that the new technologies being developed not only have greater capabilities but also bring unprecedented risks of loss of control. Once problems arise, the potential damage could far exceed that of a single nuclear power plant meltdown.

In addition, Marcus criticized the current measures taken by relevant government agencies, such as the U.S. White House administrative departments, which only require increased information disclosure. He said this is equivalent to gently asking criminals to report their criminal records monthly. He emphasized that if the government continues to delay taking stronger regulatory measures, such as mandatory temporary recalls, it will be difficult for the technology companies and even the White House to avoid responsibility once a major disaster becomes uncontrollable.