Google DeepMind chief executive Demis Hassabis is calling for a new, independent body modeled on the U.S. brokerage regulator FINRA to test the most powerful AI systems before they reach the public, laying out the proposal in a July 14 essay titled "A Framework for Frontier AI and the Dawning of a New Age."
As Hassabis describes it, the body would be a public-private partnership: funded by AI developers but governed independently, with a board that includes technical experts, open-source community representatives and government officials, and backed by federal oversight. It could contract evaluations out to specialist AI-safety organizations rather than build all of its own testing capacity, an approach Hassabis argues would let it keep pace with labs that are already moving fast.
The strength of this approach is it would be technically focused, while at the same time supporting innovation and incentivising responsible behaviour.
Demis Hassabis, chief executive, Google DeepMind
From voluntary reviews to a market gate
Hassabis wants the body running before the end of 2026. In its first phase, frontier labs would voluntarily submit models for review up to 30 days before release; participation would later become mandatory, with compliance a prerequisite for any frontier-class model to be deployed to American users. Eligibility thresholds for which models count as "frontier" would be updated as capabilities advance, and testing would concentrate on cybersecurity, biological, nuclear and other national-security-relevant risks as labs move closer to artificial general intelligence.
The proposal follows criticism of the ad hoc government reviews that preceded the release of Anthropic's Mythos and OpenAI's GPT-5.6 Sol models, which critics said lacked consistent technical expertise and a transparent decision process. It also lands as the White House has signaled it does not want a heavier-handed regulator: AI adviser Sriram Krishnan has said there will be no AI equivalent of the Food and Drug Administration, making a self-funded, FINRA-style model one of the few structures industry and the administration might both accept.
Hassabis has not said who would chair the body or how its funding would be split among labs, and rival companies have yet to publicly endorse the plan. Its prospects will likely hinge on whether Congress or the administration is willing to attach legal teeth to what remains, for now, a voluntary framework — a question made more pointed by OpenAI's disclosure this week that its own test models breached a third-party platform without human direction.