Demis Hassabis spent the weeks before his departure from Google DeepMind quietly lobbying Treasury Secretary Scott Bessent, White House technology adviser Michael Kratsios, and the leadership of rival laboratories on behalf of an independent frontier-A.I. safety body modeled on the International Atomic Energy Agency, according to a Wall Street Journal report Wednesday citing people familiar with the conversations. It’s a striking exit maneuver: the outgoing CEO of the industry’s most scientifically pedigreed lab burning his final political capital to build the regulator that’ll constrain his successor.

The public version arrived July 14, when Hassabis published a framework on X calling for a U.S.-led Frontier A.I. Standards Body. Labs would submit models to a thirty-day prerelease review, voluntarily at first, becoming mandatory if the process proved “effective.” Funding, he wrote, would “likely” come from industry. The structural analogy is FINRA, the self-regulatory body that answers to the S.E.C., not the agency itself.

Three days later, Bloomberg reported that the White House was already reviewing a parallel plan, developed with Bessent’s involvement and now sitting with Chief of Staff Susie Wiles. President Trump hasn’t yet seen it. Wiles, per Bloomberg, typically engages only when a policy debate carries political risk for the president, which tells you how the internal handicappers are reading the file.

The industry chorus was rapid and coordinated. Axios reported praise from Sam Altman, Satya Nadella, and Elon Musk; Anthropic co-founder Jack Clark called the Hassabis framework “excellent.” Altman has separately floated an IAEA-style international body that would certify countries and companies, per the Financial Times. That the four largest incumbents are aligned on the shape of their own regulator is, itself, the story.

Council on Foreign Relations analysts published a caution on July 23. Institutional form, they noted, is “never the hard part.” An industry-funded evaluator invites the conflicts that discredited the issuer-pays credit-rating model in 2008, and any body holding prerelease access to frontier weights would instantly become “one of the world’s highest-value espionage targets.” Both problems are load-bearing.

What’s on the table isn’t whether frontier A.I. gets regulated. It’s who writes the standards, who staffs the referee, and who pays the bill. Hassabis, on his way out, appears to have secured pole position on all three.

Sources