Many developers have worried that an AI model could be cut off at the government's discretion, without any legal process. In June 2026, access to Anthropic's latest models was halted immediately, on national security concerns alone, and OpenAI delayed the release of a new model at the government's request. As the industry is pressed to answer how to guard against arbitrary enforcement, Demis Hassabis, CEO of Google DeepMind, published his own proposal on July 14, 2026. At its core is the question of who holds the reins of regulation.
What Hassabis's proposed "standards body" would look like
On the morning of Tuesday, July 14, 2026, Demis Hassabis (CEO of Google DeepMind) published a personal manifesto on his Substack titled "A Framework for Frontier AI and the Dawning of a New Age." Its centerpiece is an independent "standards body" modeled on FINRA, the Financial Industry Regulatory Authority. It is not direct government control. Hassabis explains that the organization would have independent technical experts and representatives of the open-source community on its board, and could take the form of either a public-private partnership or a self-regulatory organization. The final structure has not yet been decided.
Participation would initially be voluntary. Frontier labs would share models with the standards body up to 30 days before release and undergo testing in high-risk areas such as cyberattacks, biological weapons, and "deception" (behavior in which an AI deliberately misleads humans). The framework is said to apply to all frontier-level models, "regardless of country of origin or whether they are open or closed." However, initial participation would be voluntary, and any future mandate would be limited to deployment in the U.S. market. The specific performance criteria for defining "frontier-level" would be set by the standards body itself after the fact, and the published document gives no figures.
Hassabis's own text contains no stated launch date. But in an exclusive interview, Axios reported that Hassabis hopes the standards body could be operating within this year.
What happened in June, and how the government intervened
The story begins on June 12, 2026. At 5:21 p.m. Eastern that day, the U.S. government issued an export control directive barring foreign nationals from accessing Anthropic's "Claude Mythos 5" and "Claude Fable 5." Lacking any means of verifying nationality in real time, Anthropic complied by temporarily suspending access to both models for all users. The trigger was reportedly a suspected limited jailbreak (a technique for bypassing safety measures to make a model produce inappropriate output) found in Fable 5. Anthropic pushed back, arguing that "a suspected narrow, non-generalizable jailbreak should not be grounds for recalling a commercial model deployed to hundreds of millions of people," but the suspension took effect the same day.
The episode shows a gap between the directive's nominal scope and its actual impact on the market. A measure that was legally aimed at foreign nationals took the form of a suspension for all users, because of a company's own decision when it could not tell who was who. On June 26, the U.S. Department of Commerce approved restoring access to Mythos 5 for U.S. organizations, and restrictions on Fable 5 were lifted on June 30, with global availability resuming on July 1. In other words, by the time Hassabis published his proposal on July 14, both models were already available again in some form. Also in June, OpenAI was made to limit the release of its GPT-5.6 lineup (Sol, Terra, Luna) to "a small number of trusted partners" at the government's request, with general availability slipping to July 9.
From the companies' perspective, throughout June they could not accurately predict in advance when, and to what extent, their models would be halted. Axios reported that Hassabis himself called the government's actions "a bit of a wake-up call."
FINRA or FAA: the split between Google and Anthropic
The FINRA-style approach Hassabis favors would start out voluntary, and even the line defining "frontier-level" would be drawn after the fact by the standards body itself. The board would include independent technical experts and open-source representatives, but there is no mention of allocating seats to specific companies.
By contrast, the policy document "Policy on the AI Exponential," published June 10 by Dario Amodei (CEO of Anthropic), advocates an FAA-style framework, modeled on the Federal Aviation Administration. Like type certification for aircraft, it would mandate testing and audits by third-party bodies, and the government could order the suspension or withdrawal of models that fail to meet high safety standards. The thresholds are also concrete: the rules would cover companies that develop models exceeding 10^25 FLOPs and also meet at least one of two conditions, either annual AI revenue above $500 million (about ¥81.2 billion at ¥162.41 to the dollar) or AI R&D spending above $1 billion (about ¥162.4 billion at the same rate).
The FAA-style approach sets numerical criteria in law up front, though they are not fixed. Amodei's own document leaves room to shift from compute-based to capability-based thresholds as model capabilities advance, and notes the option for regulators to start with lighter enforcement and review it in stages. Even so, the difference from the FINRA model lies in who writes the rules first. Under the FINRA model the board sets the criteria after the fact, giving industry considerable room to shape the system's design. The FAA model, by contrast, writes into law in advance mandatory testing by third-party bodies and civil penalties that scale with global annual revenue. Ironically, Amodei published his proposal just two days before Anthropic's own models were halted on national security grounds.
The FAA model's thresholds, drawn clearly by compute, revenue, and R&D spending, would impose a heavy audit burden on major companies like Anthropic, while its effect on emerging and open-source developers below those thresholds would be limited. The FINRA model, conversely, leaves the standards and board composition to be decided by the standards body afterward, so whether a company can take part in the initial design is likely to shape its future influence. FINRA has a track record as a self-regulatory organization for the U.S. financial industry under the oversight of the SEC (Securities and Exchange Commission), but experts are said to be divided on how effective a self-regulatory framework would actually be. Even with technical experts and open-source representatives on the board, it is unclear whether a framework lacking enforcement power could match the speed of the immediate government suspensions seen in June.
Why an industry facing an anti-regulation administration is asking for rules
On June 2, the Trump administration signed an executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security." Its central feature is that pre-release review of frontier models remains strictly voluntary, and the text states that it does not give the government authority to create mandatory licensing or pre-approval regimes. Sriram Krishnan, then the White House AI policy adviser (he has since left the post), said in a July 3 interview with the Financial Times: "We're not going to build an FDA for AI," and that a system requiring models to go through teams of lawyers before release "would be like throwing sand in the gears of AI progress."
Under an administration averse to regulation, the industry's top figures are themselves calling for a regulatory framework. The reason lies in what happened in June. The June 12 action was taken by the Commerce Department's Bureau of Industry and Security (BIS) under its existing export control authority. No dedicated review law for AI models had been put in place beforehand, and with the scope of application and implementing rules undefined, access to the models was cut off and then restored. What Hassabis and Amodei have in common is a desire to codify standards and procedures in advance, so that if a model is halted again, there are rules to point to. The FINRA and FAA models differ on how much authority to grant, but they agree that they are preferable to ad hoc, case-by-case intervention.
Where the contest is heading, and what it means for Japan
On June 2, the executive order set out a voluntary-review policy. On June 10, Amodei proposed the FAA model. Two days later, on June 12, the government halted Anthropic's models. On June 26, Mythos 5 was partially restored for U.S. organizations, and on June 30 restrictions on Fable 5 were lifted. On July 3, Krishnan ruled out mandatory licensing, and on July 14, Hassabis proposed the FINRA model. Arguments over the regulatory framework and actual government intervention piled on top of each other within just a month and a half.
There is no clear resolution in sight. Hassabis has outlined a path in which, if the review methods prove effective, the standards body would be quickly formalized, and passing review would become a requirement for deploying models in the U.S. market. However, specific penalties for non-compliance are not stated, so early-stage effectiveness would depend on incentives and reputation. Beyond refining review methods, including how to measure "deception," the challenge ahead is how far that effectiveness can be strengthened.
Amodei's proposal, for its part, remains for now the position of a single company, Anthropic. Which framework Congress chooses as the basis for legislation, or whether a compromise emerges, will not be clear until the coming months of debate. A situation in which AI-specific review law remains absent and enforcement is left to discretionary use of existing law could be the worst option even for companies averse to regulation. The events of June bore that out.
The issue is not irrelevant to users in Japan either. If U.S.-led certification standards become established as a de facto international standard, Japanese companies involved in developing frontier-level models may eventually be pressed to comply with similar reviews. Which of the two camps, Google or Anthropic, implements its blueprint first could change the shape of those standards themselves.