Barely a month after the Fable 5 shutdown showed the world that Washington can switch off the most powerful AI models overnight, the industry has coalesced around a plan to make sure it never happens so chaotically again. The catch, critics say, is that the plan may leave Washington’s hand on the switch, only with a smoother mechanism to pull it.
On Tuesday, Google DeepMind co-founder and CEO Demis Hassabis published a manifesto on X titled “A Framework for Frontier AI and the Dawning of a New Age”, calling for a new body to vet the world’s most advanced AI models before release. Within a day it had drawn endorsements from the chief executives of Microsoft, OpenAI and xAI, an unusual show of unity in an industry more accustomed to bare-knuckle rivalry.
At the centre of the plan is a “standards body” modelled on the Financial Industry Regulatory Authority (Finra), the private, industry-funded organisation that polices Wall Street under US government oversight. Backed by Washington but funded by the AI labs, it would test frontier models for their most dangerous capabilities – cyber, biological and “deception” or agentic risks – under a majority-independent board of technical experts, with industry, government and open-source representatives alongside.
“Initially, frontier labs would voluntarily share models with the standards body for review up to 30 days before release,” Hassabis wrote. “Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that frontier models would be required to pass it to be deployed in the US market.”
The rules would apply to any frontier-class model “no matter their country of origin or whether they are open or closed”, with start-ups and academic systems exempt. Hassabis expects the “frontier” designation itself to incentivise participation, telling Axios that being tested signals a lab matters: “I think that’s a pretty nice, prestige kind of asset to have.”
Hassabis, who spent months briefing the Trump administration, fellow lab leaders and European officials before going public, told Axios he wants the body operational “before year-end”, arguing that artificial general intelligence is “probably only a few short years away”. The idea builds on the pitch AI chiefs made to world leaders at the G7 summit in Évian in June.
Industry closes ranks
Microsoft CEO Satya Nadella called it “an important piece from Demis”, adding that the goal is an ecosystem “that promotes innovation and choice, while avoiding any one model drop that breaks the world”. OpenAI’s Sam Altman signalled support, having used a Financial Times op-ed earlier this month to call for a US-led international forum along similar lines.
Even Elon Musk, whose xAI is the only major frontier lab outside the industry’s Frontier Model Forum safety group, offered a measured endorsement. “It is a thoughtful framework overall and certainly a good starting point for discussions,” he wrote on X on Wednesday – a shift in tone from June, when he greeted the push for an AI regulator by jokingly proposing the name “AIAIAI”, pronounced “ay yai yai”.
Read: The Popia problem with agentic AI
Anthropic’s Dario Amodei agrees with the direction but not the design: in a June essay he argued for an FAA-style federal agency with statutory power to block unsafe models. As Axios put it, the lab chiefs now agree Washington should regulate them, “differing mainly on who holds the authority”.
The Hassabis proposal’s clearest gap concerns China. Hassabis wants his standards applied to every frontier model regardless of origin, but does not spell out how a US-anchored body would compel Chinese developers – such as DeepSeek and Zhipu AI, which are closing in on frontier-class capabilities – to submit models for testing.
Beijing’s leading labs are in any case pursuing the opposite strategy: Zhipu founder Tang Jie argues that security comes from broad access rather than technological barriers, and released his latest model as open source. Once a model’s weights are published, no authority anywhere can recall or pause it.
The Fable 5 test
The proposal’s other open question is whether it would have prevented the crisis that inspired it. In June, US commerce secretary Howard Lutnick placed Claude Fable 5 and Mythos 5 under export controls days after launch – stretching a legal regime built for physical goods to a cloud service. This after Amazon researchers jailbroke a Fable 5 safeguard by, in Anthropic’s account, simply asking the model to read a codebase and fix its flaws, whereupon it identified vulnerabilities and in one case generated exploit code.
Anthropic disputed the severity, showing less capable rival models could do the same, but was forced to disable both models worldwide overnight, restoring access only two weeks later after a negotiation it said had no rules, protocol or playbook.
Read: AI Barometer: the best AI for every job right now
Under Hassabis’s regime, Fable 5’s capabilities would have been tested against a published protocol before release, and post-release flaws handled through a managed patching process rather than a Friday night directive. The promise is predictability. What it would not do is remove Washington’s underlying export-control powers – a determined administration could still act outside the protocol, as it did in June – and history suggests voluntary regimes have limits.
Brookings fellow and former US Federal Communications Commission chair Tom Wheeler has noted that comparable international standards bodies, from the Financial Action Task Force to the Basel banking committee, worked because non-compliance carried real penalties. “Voluntary compliance means the outcome is only a suggestion, not a standard,” he said.
He also warns that a process limited to the companies “solves the companies’ problems while giving governments cover to appear to be doing something”.
Nor is it clear what would stop the body becoming an instrument of the Trump administration. Finra-style independence is the design answer, but the body would still be, in Axios’s phrasing, “answerable to the US government” – and the White House has been cool to formal regulation, with Sriram Krishnan, its senior AI adviser until end-June, telling the Financial Times “there will not be an FDA for AI”, referring to the US Food and Drug Administration.
‘A standing policy tool’
The R Street Institute warned after the shutdown that the export-control order had become “a standing policy tool” deployable “against any lab, model or company that becomes inconvenient”. A standards body layered on top does not retract it.
For the developing world, frontier access would still be granted, not guaranteed, by a gatekeeper in Washington. Nothing in the framework gives South Africa or any other African government a voice in the body that would decide which models reach their markets – the board Hassabis envisages draws from industry, the US government and the open-source community. For countries that consume frontier AI but do not build it, a smoother gate is still a gate.
What happens next
Hassabis wants the body running within months. But enforcement, board composition beyond the industry that funds it and China’s participation all remain unresolved. Until then, the lesson of June stands: whoever controls the deployment gate controls the world’s access to frontier AI, and that gate remains in Washington. – © 2026 NewsCentral Media
