Google DeepMind CEO Demis Hassabis has published an unusually concrete proposal for a US-led Frontier AI Standards Body modelled on the Financial Industry Regulatory Authority (FINRA), warning that artificial general intelligence — systems with human-level cognitive capabilities — could arrive “probably only a few short years away.”

The proposal, published on Hassabis's personal Substack on July 14 and simultaneously discussed in a The Economist interview, goes further than any major AI lab CEO has previously put in writing. It calls for an industry-funded, independent body that would review the most advanced AI models up to 30 days before release, with a pathway from voluntary compliance to mandatory pre-clearance, and even a mechanism to “coordinate a slowdown” among frontier labs if safety concerns demand it.

The intervention arrives at a moment when the US government has already begun exerting ad hoc pressure on frontier AI developers. President Donald Trump's Executive Order 14409 on June 2 created a voluntary 30-day government-review framework, and subsequent actions saw Anthropic's Mythos and Fable 5 models restricted by export control order on June 12 while OpenAI limited GPT-5.6's release to government-cleared partners on June 26.

What Hassabis Proposed: A FINRA-Style AI Watchdog

Hassabis's framework envisions a public-private partnership that would operate independently of both government and the industry it regulates. The body would be funded by the frontier labs themselves but governed by a board of independent experts — including representatives from open-source communities and civil society — and would maintain a technical testing regime updated quarterly to keep pace with rapid capability advances.

“Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release,” Hassabis wrote. If the framework proves effective, that review could become “required to pass it to be deployed in the US market.”

The proposed body would test frontier models across a range of catastrophic-risk domains: cybersecurity vulnerabilities, biological threat potential, nuclear risks, deception capabilities, and autonomous system risks including the ability to bypass guardrails and engage in “long-term, agentic, deceptive behaviour.” Non-frontier models — those from startups or academic researchers below a capability threshold — would be entirely exempt from the process, a deliberate carve-out intended to avoid stifling innovation.

Notably, the framework includes an escalation clause that Hassabis said “could be ratcheted up if the seriousness of the situation demands, including coordinating a slowdown in development among the Frontier Labs if deemed necessary.” This provision for a collective pause across leading labs goes beyond what any major AI CEO has previously endorsed in writing.

The choice of FINRA as a model is deliberate. Just as FINRA oversees broker-dealers in the US securities industry — setting standards, conducting examinations, and enforcing compliance — Hassabis's proposed body would operate at arm's length from both government and corporate influence while being funded by the entities it regulates. Publishing the blueprint on a personal Substack rather than a corporate blog, paired with a high-profile Economist interview, signals a deliberate public-policy intervention aimed at shaping the debate rather than a routine thought leadership piece.

AGI Timeline Accelerates

Perhaps the most arresting claim in Hassabis's post is his accelerated timeline for AGI. He now places it at “probably only a few short years away” — a significant tightening from his estimate at the India AI Impact Summit in New Delhi just five months earlier, where he told 250,000 attendees that AGI was “on the horizon, maybe in the next five to eight years.”

Hassabis did not explicitly address why his timeline has compressed. The shift could reflect accelerating capability advances inside DeepMind and other frontier labs, or it could be a rhetorical choice intended to lend urgency to his regulatory proposal. Either way, it places the governance debate in a new light: if AGI is genuinely only a few years off, the window for establishing effective oversight is considerably narrower than previously assumed.

Hassabis described the potential impact of AGI in sweeping terms, writing that it could deliver “10 times that of the Industrial Revolution at 10 times the speed.” The comparison, speculative but striking, underscores why he believes institutional scaffolding for safe development must be built now.

The Precedent: Governments Already Acting

Hassabis's proposal arrives against a backdrop of unprecedented — and controversial — government intervention in AI. Trump's Executive Order 14409, signed June 2, established a voluntary framework under which “covered frontier models” can be submitted to the government for 30-day pre-release review. The order explicitly states it does “not authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement.”

Yet in practice, the line between voluntary and mandatory has already blurred. On June 12, Anthropic's Mythos and Fable 5 models were effectively disabled in the US following an export control order — a move that surprised the industry and raised questions about whether executive authority was being used to impose de facto licensing. The incident sent ripples across the global AI industry. In India, the Nasscom chairman subsequently called for a tenfold increase in AI investment, arguing the episode exposed systemic vulnerabilities in how frontier models interact with enterprise environments. Nasscom Chief: 10x AI Spend Needed After Mythos Sandbox Scare

OpenAI followed on June 26, agreeing to limit GPT-5.6 — released in Sol, Terra, and Luna variants — to a small group of government-cleared partners after a request from Washington. The Indian Express, which broke the story of Hassabis's proposal for Indian audiences, explicitly framed it against these events, noting that “the US government has already exerted ad hoc pressure” on both Anthropic and OpenAI. TechCrunch connected the same dots, reporting that the framework is “designed to replace the current patchwork of emergency measures with a predictable, technically competent oversight system.”

Critics on Hacker News and elsewhere have argued that Hassabis's proposal — coming from the CEO of one of the world's most advanced AI labs — would create a regulatory moat that benefits incumbent frontier developers at the expense of open-source projects and international competitors. Hassabis counters that the exemption for non-frontier models addresses this concern, and that the urgency of catastrophic risk justifies institutional guardrails.

What This Means for India

Hassabis's proposal carries particular significance for India, where he was a featured speaker at the India AI Impact Summit in New Delhi just five months ago. There, Prime Minister Narendra Modi inaugurated the gathering of over 250,000 attendees at Bharat Mandapam, and Hassabis called India a future AI “powerhouse” while announcing expanded Google DeepMind research partnerships with Indian institutions. India's startup Sarvam AI also launched its Sarvam-30B and Sarvam-105B LLMs at the same summit.

The proposed US-led Frontier AI Standards Body would have direct implications for India's rapidly growing AI ecosystem. If the framework becomes mandatory for deployment in the US market, Indian AI developers — including startups and enterprise players — would need to navigate a new regulatory landscape for cross-border AI products. Conversely, the exemption for non-frontier models could benefit India's startup ecosystem, most of which operates well below the frontier-class capability threshold. India's ambition to build semiconductor self-reliance is also intertwined with the responsible AI debate, as chip design increasingly incorporates AI-driven verification and security testing that could fall under emerging standards. Can Responsible AI Catalyse India's Semiconductor Self-Reliance?

India's own cybersecurity landscape underscores the relevance of frontier AI safety. According to government data, the country recorded 22.68 lakh cybersecurity incidents in 2024, more than double the 10.29 lakh recorded in 2022. As AI capabilities advance, the potential for AI-powered cyber threats targeting Indian infrastructure, financial systems, and government networks grows correspondingly. A global standards framework could help establish defensive norms that benefit countries like India.

There is also a strategic dimension: India aspires to be a participant in shaping global AI standards, not merely a rule-taker. Hassabis's framework envisions eventual international adoption — he says it “could become the basis for international standards” — but for now remains explicitly US-led. How India engages with this process, whether through bilateral AI dialogue with the US or through its own domestic frameworks, will shape the country's position in the emerging global AI governance order.

Open Questions

Hassabis's post leaves several critical questions unanswered. Whether the Trump administration or Congress will adopt his specific institutional design is unknown — the White House has not commented on the proposal. The relationship between his proposed body and the existing AI Safety Institute established under EO 14409 remains unclear. International adoption beyond the US is aspirational only. And crucially, how “frontier-class” capability thresholds would be set and who controls them is yet to be defined.

By publishing a concrete institutional blueprint rather than a general call for safety — and by accelerating his own AGI timeline — Hassabis has shifted the terms of the debate. The question is no longer whether frontier AI should be overseen, but who should do it, under what rules, and whether the world can build the institutions fast enough.

Sources