Google DeepMind CEO Demis Hassabis published a framework manifesto on his personal Substack in July, arguing for the creation of an international regulatory body independent of governments and corporations, modeled after the U.S. Financial Industry Regulatory Authority (FINRA) framework. It calls for all AI models that meet the frontier threshold to undergo mandatory review for up to 30 days before release.
The core design of Hassabis’ manifesto
Hassabis chose FINRA as the reference framework: a self-regulatory organization funded by the industry, operating independently and supervised by the U.S. Securities and Exchange Commission (SEC). It audits financial institutions’ compliance and has enforcement powers to issue fines and impose suspensions. He wants to apply the same logic to AI regulation—funded by the industry, but with decision-making authority granted to independent figures who are not affiliated with any single company.
The main specifications for this regulatory body are as follows:
Board composition: Most seats reserved for independent technical experts at the Turing Award level, with additional seats allocated to representatives from industry, government, and open-source communities
Funding source: Fully provided by profitable top AI labs
Review targets: All models that reach the frontier threshold, whether open-source or closed-source, regardless of where a company’s headquarters is located
Review period: Up to 30 days, with the threshold itself updated regularly as models’ capabilities evolve
Expansion mechanism: The intensity can be increased depending on the severity of the situation; in the most extreme cases, it may coordinate multiple major labs to slow down R&D efforts together
Target timeline: Hassabis hopes the institution begins operating no later than the end of 2026
The trigger for publishing the proposal early
Hassabis had been developing this framework for months, but what truly prompted him to publish it early was Anthropic’s Mythos model, which was limited-release in April 2026. After the model debuted, it demonstrated cybersecurity offense-and-defense capabilities far beyond expectations. This sparked concern in both the industry and policy circles about a “new era of AI-driven cybercrime.” Hassabis directly labeled it a “warning sign,” saying it was no longer acceptable to wait for regulators to catch up slowly.
In an interview, Hassabis also mentioned longer-term biosecurity risks—specifically, the ability of AI models to help design pathogens or biological weapons and other threats to public health. He said: “The models aren’t strong enough to that degree yet, but if we extrapolate from current progress, we could see it within a few years.”
At the same time, under pressure from Trump’s government, both Anthropic and OpenAI delayed large-scale releases of their latest products. This led to negotiations between the European Union and the U.S. government over model access rights, and multiple AI company CEOs also met with leaders from their respective countries during the French G7 summit to discuss the issue.
Statements from Altman, Musk, and Amodei
Before the framework was publicly announced, Hassabis privately showed the draft to nearly all major AI labs, including Anthropic CEO Dario Amodei. After the public release, OpenAI CEO Sam Altman called the idea “very thorough.” Even Elon Musk, who typically takes a stance opposed to Altman, said it was “a very good starting point,” and the two unexpectedly aligned on the same position.
Amodei had earlier suggested that the U.S. establish a dedicated regulator similar to the Federal Aviation Administration (FAA) to oversee AI, and Altman had also proposed a similar international regulatory initiative. When Hassabis spoke with Bloomberg, he said: “I think this is highly compatible with the ideas Dario and Sam have been putting forward all along. We just need a concrete plan, not an abstract concept.”
Implementation challenges: U.S. legislative gaps, Trump’s shifting stance, and divergence among the regulatory paths of three parties
An AI industry leader who requested anonymity said that industry consensus alone isn’t enough—the real difficulty is political. The U.S. Congress has yet to pass any meaningful federal AI legislation. Trump’s government’s stance toward AI has been wavering between permissiveness and intervention. And the regulatory paths of the U.S., the EU, and China often diverge.
Hassabis plans to travel to Washington next week to discuss directly with U.S. policymakers. He said he is optimistic that the current momentum may be enough to carry the effort to the finish line, but he also knows: “What comes next is figuring out how to actually implement it, not just publish a blog post.”
FAQ
How is the AI regulatory body proposed by Hassabis similar to FINRA?
FINRA is a self-regulatory organization funded by participants in the U.S. securities industry, operating independently and supervised by the SEC. It audits financial institutions’ compliance and has enforcement powers. Hassabis’ proposal adopts the same logic: funded by profitable AI labs, led by independent technical experts who are not affiliated with any government or company, and carrying out mandatory pre-release reviews for frontier AI models.
Why is Anthropic’s Mythos model seen as the trigger for the proposal?
After Anthropic’s Mythos model debuted in April 2026, it demonstrated cybersecurity offense-and-defense capabilities far beyond expectations, raising concerns in the industry and among policymakers about “AI-driven cybercrime.” Hassabis labeled it a “warning sign,” saying the incident led him to publish the regulatory framework he had been developing for several months earlier, rather than continuing to wait.
When might this international AI regulatory body begin operating?
Hassabis said he hopes the institution can begin operating as early as the end of 2026, but the timeline faces major political obstacles: the U.S. Congress has not passed any federal AI legislation to date; Trump’s government’s position is unstable; and there are also differences in regulatory approaches among the U.S., the EU, and China. Hassabis’ trip to Washington next week is the next concrete step, with an official announcement serving as the reference.
