Demis Hassabis wants the United States to build a Wall Street style referee for artificial intelligence, and he wants it running before the year ends.
The Google DeepMind chief executive has put a deadline on a fight Washington keeps trying to manage case by case. In a manifesto published Tuesday titled "A Framework for Frontier AI and the Dawning of a New Age," and in an exclusive interview with Axios, Hassabis called for an independent standards body modeled on FINRA, the Financial Industry Regulatory Authority that polices Wall Street under the SEC's oversight.
The pitch is straightforward. Industry pays for it. The government recognizes it. The world's most powerful AI labs hand over their models before release so outside reviewers can test whether they are dangerous enough to slow down.
Under Hassabis's plan, frontier labs would share models up to 30 days before launch. Reviewers would probe for dangerous capabilities in cybersecurity, biological risks and what Hassabis calls "deception," meaning a model's capacity to mislead its own operators. Participation would start voluntary. But Hassabis wrote that once the testing regime proves "effective and robust," formalization "could quickly follow." At that point, passing the review could become a condition of deploying a frontier model in the U.S. market at all.
He wants a board where independent members hold the majority, with Turing Award winners and other credentialed technical experts sitting alongside industry, government and open source representatives. That's a heavy bar for legitimacy. He also wants the rules to apply evenly. Axios reported that the standards would cover every frontier-class model "no matter their country of origin or whether they are open or closed."
Washington made room for this #
Timing matters here. The Washington Post reported in May that President Trump declined to sign an expected AI executive order after last-minute pressure from Silicon Valley figures, including former AI and crypto adviser David Sacks, who warned that early government review could become a de facto approval system. The draft would have let companies give the government access to frontier models up to 90 days before public release.
Then Washington moved anyway. The Associated Press reported that Trump signed a June 2 executive order creating a voluntary framework for federal review of the most advanced AI systems for up to 30 days before public release. That is not the same as a formal licensing regime. But don't pretend it is nothing. When the government can influence who gets early access to models and when foreign users can touch them, the line between voluntary review and practical permission starts to blur.
That is exactly the opening Hassabis is stepping into. A FINRA-style body, funded by industry and answerable to government, gives Washington something sturdier than improvisation without creating a new agency that looks like an FDA for AI. You can see why that appeals to a White House trying to hold two ideas at once: beat China on AI, and stop the next frontier model from becoming a cyber weapon before anyone has read the manual.
The case for a standing review process is no longer abstract. AP reported last week that the Trump administration lifted restrictions on Anthropic's Claude models after a cybersecurity alarm, ending a weekslong ban tied to federal concerns. Anthropic said the Commerce Department had blocked foreign nationals from using Claude Fable 5 and Mythos 5 on June 12, forcing the company to take the products down for all users just days after unveiling them.
The trigger, according to Anthropic's own account cited by AP, was a report from Amazon researchers who found a way to bypass Fable 5's safeguards and use it to discover and potentially exploit software vulnerabilities. Mythos 5 remains available only to select U.S.-based organizations approved by the federal government. OpenAI also restricted GPT-5.6 Sol to government-approved customers for a temporary period, AP reported.
That is messy governance. It may even be necessary, but it is still messy. Companies do not know exactly what trips a delay. Officials do not have a permanent expert institution built for this job. Users find out after the fact that a model they expected to test or buy has been pulled into a national security process.
The trust problem is the real test #
Whether rivals sign on is the hard question. A body proposed by Google's own AI lab chief, with a board that industry helps staff, is not obviously neutral ground for OpenAI, Anthropic, xAI or Meta. Each has its own safety framework and its own reason to prefer setting the terms itself. Trust is not automatic.
Frankly, the FINRA comparison only works if the body gets real funding, real compute access and real submission commitments from labs that have spent the past year racing each other on release timelines. A 30-day pre-release window is a real cost when your rival might ship in half that time. That is not a philosophical objection. It is how product competition works.
Still, Hassabis has chosen a practical structure. He is not asking Congress to pass a sweeping AI safety law in a month. He is asking the labs to build a referee before the referee is built for them. That is a different kind of pressure, and it is aimed directly at chief executives who have spent years saying they welcome oversight while fighting over who should write the rules.
For investors pricing AI valuations, the proposal is a signal worth watching rather than a settled outcome. If frontier labs agree to submit models for review, and if that review becomes a de facto gate to the U.S. market, release speed becomes a financial variable. Revenue timelines change. So do model launch strategies.
If the proposal stalls, the current system continues: voluntary federal review, sudden restrictions, negotiated access and public confusion over what the government has actually approved. Hassabis wants the watchdog operating before year-end. That gives him less than six months to prove the industry can regulate itself before Washington decides how much self-regulation is really worth.
Also read: Mitsubishi Bets 7.5 Billion Dollars That Gas Power Is AI's Real Bottleneck • Dave Clark Bets His Second Act on AI Agents Running Supply Chains • OpenAI Researcher Miles Wang Wants $200 Million to Bet on Failed Drugs