Google DeepMind CEO Demis Hassabis is calling on the United States to establish a federally overseen standards body that would test frontier artificial intelligence models for national security risks before public release. The proposal, published July 14 in a Substack essay titled “A Framework for Frontier AI and the Dawning of a New Age,” arrives as global AI governance splinters across competing regulatory frameworks and an independent safety index finds no major AI lab scoring above a C+.
What Is Hassabis Proposing and How Would It Work?
Demis Hassabis outlined a public-private partnership modeled after the Financial Industry Regulatory Authority (FINRA), the self-regulatory organization that oversees U.S. broker-dealers. Under the framework, AI developers whose models meet certain capability thresholds would be designated as “Frontier Labs” and encouraged to publish model cards, maintain cybersecurity standards, and vet personnel handling advanced systems.
The proposed body would begin with a voluntary phase in which frontier labs submit models for independent review up to 30 days before release. Evaluators would test for dangerous capabilities across cybersecurity, biology research, and deceptive behavior. Demis Hassabis wrote that specific tests “could look for attempts to bypass safety guardrails or signs of deception, and ensure best practices, such as digitally watermarking AI-generated images and generating human-readable output tokens to understand model reasoning.”
Once the evaluation protocol proves effective, Demis Hassabis envisions it becoming mandatory. Models judged too risky could trigger a coordinated slowdown among participating labs. The private sector would fund and supply the bulk of the necessary computing infrastructure, while an independent board of technical experts and open-source representatives would direct the work.
Why Is the Proposal Coming Now?
The essay frames artificial general intelligence as “probably only a few short years away” and describes the current moment as “the foothills of the singularity.” Demis Hassabis compared the potential impact to “10x of the Industrial Revolution at 10x the speed,” arguing that commercially intense and geopolitically charged competition is outpacing the field’s understanding of what these systems can do.
The timing aligns with several converging pressures. A June White House executive order already sought voluntary pre-release access to frontier models for up to 30 days to test advanced cyber capabilities. Google DeepMind, Microsoft, and xAI previously agreed to provide models for federal national security testing. Separately, tensions escalated after the U.S. government restricted access to certain advanced AI systems over concerns that safety controls could be bypassed, raising questions about whether ad hoc reviews are sufficient for increasingly capable technology.
Demis Hassabis and Anthropic CEO Dario Amodei reportedly made a similar case at a G7 meeting in June, calling for a U.S.-led coalition to shape AI rules in conversation with heads of state that included President Donald Trump.
What Does the AI Safety Index Reveal About Industry Readiness?
The proposal landed the same week the Future of Life Institute published its Summer 2026 AI Safety Index, and the results reinforce the urgency behind it. The index evaluated nine leading AI companies across 37 indicators spanning risk assessment, safety frameworks, governance, and transparency. An independent panel of seven researchers and governance experts assigned grades using the U.S. GPA system.
Anthropic earned the highest overall grade at C+, leading five of six evaluation domains. OpenAI received a C, with particular strength in risk assessment. Google DeepMind also scored a C. Meta improved from sixth to fourth place with a D+, while xAI dropped from fourth to seventh with a failing grade. DeepSeek and Mistral also received F grades, making inadequate safety a global problem spanning U.S., Chinese, and European developers alike.
The most striking finding was not any single company’s grade but a pattern across the industry. The panel found that Anthropic, OpenAI, Google DeepMind, and Meta have all weakened or voided earlier pledges to pause development if their systems approached certain danger thresholds. UC Berkeley professor Stuart Russell, one of the panelists, noted that “the capabilities race has become more extreme” and that companies are now “planning to release them even if it’s demonstrably unsafe to do so.”
How Is Europe Approaching the Same Problem?
The European Commission is building its own answer. On July 7, the Commission published an Action Plan on Cybersecurity and AI that establishes pre-market evaluation of advanced AI models as EU policy. The plan directs the Commission and the European Union Agency for Cybersecurity (ENISA) to develop a “European Blueprint” for structured access to frontier AI capabilities and to create a secure testing platform for organizations in critical sectors including energy, finance, health, and transport.
The EU evaluation capacity is targeted to become operational by 2027, working in support of the European AI Office’s enforcement function. General-purpose AI transparency obligations under the AI Act take effect on August 2, 2026, adding regulatory teeth behind the testing infrastructure.
The European approach differs from the Hassabis proposal in a fundamental way. Where Demis Hassabis envisions industry-funded self-regulation overseen by government, the EU is building state-run evaluation infrastructure backed by existing legislation. Both frameworks aim to catch dangerous capabilities before deployment, but they reflect different assumptions about who should bear the cost and who holds ultimate authority.
The gap between voluntary industry pledges and binding oversight is narrowing on both sides of the Atlantic. Whether the U.S. adopts a FINRA-style model, the EU operationalizes its testing platform, or both proceed in parallel, the direction is consistent: pre-market scrutiny of frontier AI is shifting from aspiration to infrastructure. The question is whether the infrastructure can scale as fast as the models themselves.
