Nearly 400 scientists have demanded the creation of an international organization to protect and monitor advanced AI systems. Researchers including Geoffrey Hinton and Yoshua Bengio called for stricter security standards and incident reporting mechanisms to be implemented for cyber-enabled AI models.
Tech: Nearly 400 scientists have demanded the formation of an international monitoring body to ensure the security of advanced Artificial Intelligence (AI) systems. The signatories from 34 countries also include AI experts Geoffrey Hinton and Yoshua Bengio. Scientists have called for strict security standards, independent monitoring and reporting mechanisms for serious security incidents for AI models with cyber capabilities.
Demand to create an international organization to monitor advanced AI
About 400 scientists have demanded the formation of an international organization for the security and monitoring of advanced Artificial Intelligence (AI) systems. In an open letter released Tuesday, the researchers said shared security standards should be created for the development and testing of advanced AI models with cyber capabilities. The signatories also include leading experts in the AI field, Geoffrey Hinton and Yoshua Bengio.
The scientists propose that the new body set safety standards at an industry level and maintain a register of serious safety incidents. According to him, this register should also be made public over time, to ensure greater transparency about risks and incidents associated with advanced AI systems.
Emphasis on strict security standards and independent monitoring for AI models
In the open letter, the researchers suggested that state-of-the-art AI models should be kept isolated from external networks in the early stages. Along with this, there should also be independent monitoring of their security arrangements. Scientists argue that voluntary security commitments from companies alone are not enough and that more systematic arrangements are needed to manage the risks associated with advanced AI.
The signatories said entities developing or testing AI models with cyber capabilities should immediately adopt high-level security and control standards. He believes that companies should not delay taking security measures until governments are working towards developing formal monitoring systems.
The demand comes as some researchers have raised concerns about shortcomings in the testing and control of advanced AI systems. According to him, in some cases AI models crossed the boundaries of a sufficiently secure testing environment and reached external systems. Such incidents have raised questions about existing arrangements to control AI models and prevent their possible misuse.
Growing discussion on AI security in America, Britain and Europe
Different approaches are emerging regarding AI regulation at the international level. Earlier, Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a White House agreement that included a commitment to adopt voluntary security measures. However, the agreement did not contain a formal enforcement mechanism.
EU technology chief Hanna Virkkunen told Reuters on Friday that the EU's AI Act is sufficient to deal with the risks associated with uncontrolled AI systems. On the other hand, on Sunday, the Trump administration lifted the ban on giving American AI laboratories early access to models for security testing with Britain's AI Security Institute.
Different opinions about risks among AI experts
Geoffrey Hinton and Yoshua Bengio were awarded the Turing Award in the year 2018. He shared this award with Meta's AI researcher Yann LeCun. However, Lacan did not sign this open letter.
Lacan has been of the opinion that recent AI safety incidents reflect primarily engineering failures, and not necessarily existential risks associated with AI. This difference suggests that there are differing views among experts on AI security about the nature of the risks and how to deal with them.
The main objective of this initiative of scientists is to institutionalize shared security standards, independent monitoring and reporting of incidents for advanced AI systems. How the composition of the proposed international organization and its powers will be determined will further depend on the decisions of governments and relevant organizations.