Demis Hassabis calls for US-led frontier AI standards body with 30-day pre-release model testing

Demis Hassabis Calls for US-Led Frontier AI Standards Body with 30-Day Pre-Release Testing

Demis Hassabis, the Nobel laureate who leads Google DeepMind, is proposing the creation of a US-led standards body to test frontier AI models before release, citing national security risks from cybersecurity threats, nuclear, and biological dangers as capabilities advance rapidly.

The proposal follows the White House's recent export ban on Anthropic's most advanced models and fresh warnings about AI's potential to disrupt the global economy. What sets Hassabis's intervention apart is its specificity: he's offering concrete mechanisms rather than abstract calls for regulation.

The core framework

Hassabis proposes a federally overseen body—potentially an expanded version of the Commerce Department's existing Center for AI Safety and Innovation—funded by AI companies themselves. The centerpiece is a 30-day pre-release testing window where frontier models must undergo evaluation for cybersecurity, biological, nuclear, deception, autonomy, and guardrail capabilities.

Labs would submit models before launch. Smaller models and standard AI applications—Netflix recommendation systems, for instance—would be exempted, allowing them to ship without friction. National labs, federal agencies, and independent third-party auditors would conduct confidential tests designed so labs cannot train specifically against the benchmarks. Labs would be required to fix serious vulnerabilities discovered after release.

The open source and foreign model problem

Hassabis's proposal to apply rules to foreign and open source models deployed in the US creates immediate practical and political friction. The mechanism is unclear—whether enforcement happens through GPU access, cloud infrastructure, or takedown notices to platforms like Hugging Face and GitHub. The hosts acknowledge this will be "very controversial to the open source fans."

There's also a competitive concern that cuts both ways. Regulation could slow frontier closed-source labs more than nimble open source projects, giving the latter an advantage. Or it could harm open source by imposing regulatory costs that only well-capitalized companies can absorb—mirroring how FDA drug approval timelines favor large pharma over startups.

What's actually missing

The hosts push back on a recurring weakness in AI policy proposals: they lack concrete, testable scenarios. Hassabis calls for "urgent action" but doesn't spell out what triggers intervention or what happens next. One host notes that more useful proposals would say things like "if unemployment exceeds 10 percent, deploy stimulus" rather than abstract warnings about job displacement.

Hassabis also doesn't specify whether the framework would allow coordinated slowdowns among frontier labs if testing reveals serious risks—a point he mentions but leaves vague.

One host observes that Hassabis could have simply called for beefing up the existing Casey framework with specific policies rather than proposing an entirely new body. The practical question of what actual effect this would have on the industry remains unclear.