Google DeepMind CEO Demis Hassabis calls for the United States to lead an independent standards body that tests the most advanced “frontier” AI models for national security risks before they are released. In reporting summarized by multiple outlets, Hassabis proposes that such a body would cover frontier models developed in the U.S. and abroad and apply to both open and closed systems. He argues that current capabilities and trends in frontier models create urgent risks, citing cybersecurity challenges and warning that other threats could emerge as systems improve, including risks connected to biological and nuclear threats. Hassabis also frames the effort as part of managing dangers associated with advanced AI, including the potential trajectory toward artificial general intelligence.
The proposal described in the articles would resemble a federally overseen public-private arrangement, initially voluntary and later mandatory for deployment in the U.S. market once models are shown to be effective in meeting agreed testing and safety criteria. Frontier labs would initially share models for review ahead of release, and the body would conduct assessments that could include checks for attempts to bypass safety controls or deception, as well as operational best practices such as watermarking AI-generated media. The articles also say substantial funding would be required to support large-scale testing and attract technical experts.
