DeepMind CEO Proposes International AI Safety Body

Google DeepMind CEO Demis Hassabis has put forward a detailed framework for creating an international oversight body that would conduct mandatory safety reviews of advanced AI systems before they reach the public.

The proposal, published in a blog post last week, calls for independent technical experts to perform "rigorous" testing of frontier models, with companies funding the effort and submitting their systems for up to 30 days of review. The framework drew immediate endorsements from OpenAI's Sam Altman, who termed it "thoughtful," and Elon Musk, who called it "a good starting point for discussions."

Hassabis plans to present the proposal to U.S. policymakers in Washington meetings next week, according to a person familiar with the matter, as first reported by the Los Angeles Times.

Why it matters

The timing reflects growing friction between AI companies and governments over model releases. Both Anthropic and OpenAI recently delayed launches under Trump administration pressure, creating confusion among international partners and prompting emergency discussions between the EU and U.S. officials. An industry-backed oversight mechanism could provide a more predictable alternative to ad hoc government interventions that have caught even major labs off guard.

Catalyst for action

Hassabis said he accelerated his timeline after Anthropic released its Mythos model, which demonstrated advanced cybersecurity capabilities he described as "a warning shot for society." While current models don't yet pose severe biosecurity threats, he warned that extrapolating current progress suggests such risks could materialize within two years.

"I'm worried about different types of risks also coming down the line," Hassabis said in an interview, pointing specifically to potential biosecurity dangers.

Google DeepMind has already been coordinating with government officials and AI security institutes as it develops new models, Hassabis confirmed. But he argued that continuing these conversations "in an ad hoc fashion isn't very sustainable in the longer term."

Industry alignment

Before going public, Hassabis socialized the plan with virtually every major AI lab and multiple CEOs, including Anthropic's Dario Amodei, who recently proposed a Federal Aviation Administration-style agency for AI oversight in the U.S.

"I think it's very compatible with the ideas he's been putting out, and also Sam [Altman] too," Hassabis said. "But we needed something concrete rather than abstract."

The proposal models itself after existing regulatory bodies like the Financial Industry Regulatory Authority and includes provisions to "ratchet up" oversight if risks escalate, potentially coordinating development slowdowns across leading labs.

Political hurdles remain

Despite industry support, significant obstacles persist. The U.S., EU, and China have pursued divergent regulatory approaches, while Congress has yet to pass federal AI legislation. The Trump administration has oscillated between hands-off and interventionist stances.

One unnamed AI leader acknowledged that industry consensus alone won't suffice, noting the political challenges of securing government backing.

Still, Hassabis expressed optimism about current momentum. As a Nobel laureate with decades in AI research and fewer public feuds than some peers, he may be uniquely positioned to build coalitions.

"Obviously the next step is to see if we can actually make it happen," Hassabis said, "not just publish a blog post."

These details were first reported by Shirin Ghaffary for Bloomberg in the Los Angeles Times.