DeepMind CEO proposes AI standards body to test frontier models

DeepMind chief executive Demis Hassabis proposed a U.S.-led AI standards body to screen frontier models for national security risks, pushing the industry toward a regime that looks less like voluntary ethics and more like financial supervision. The model he is invoking is FINRA, the private self-regulatory body that oversees U.S. broker-dealers under Securities and Exchange Commission supervision.

Hassabis’s framework would begin with voluntary submissions from frontier labs up to 30 days before release, giving independent technical experts time to test the most advanced systems for cyber risks, biological threats, deception and other security concerns. The proposed body would be industry-funded but answerable to the U.S. government, and it would include open-source representatives alongside technical specialists. The plan would also require watermarking for AI-generated images and human-readable reasoning traces for models, two measures meant to make machine output easier to identify and audit.

The governance question is who would control the gatekeeping once the body exists. Hassabis wants the standards to be revised often and, over time, hidden from labs so companies cannot train directly to the test, a move designed to limit gaming but one that would also concentrate power in the hands of whoever writes the benchmarks. In his framework, approval could eventually become mandatory for deployment in the U.S. market, turning a voluntary review into a condition for access to the country’s largest commercial AI market. He has also said the body could coordinate an industry-wide slowdown if risks become severe.

That is where the comparison with Wall Street becomes more than a metaphor. FINRA’s authority works because broker-dealers operate in a market already shaped by longstanding SEC oversight; an AI body would have to define the tests, police compliance and handle disputes when a safety rule cuts into commercial advantage. The hardest calls would fall on benchmarks that can decide which model ships, which model waits and which company gets to claim leadership.

Hassabis has been briefing U.S. officials, AI executives and European policymakers as he presses for the new system, and reports say the Trump administration has responded positively. The pitch lands at a moment when the major labs agree some new approach is needed, even if OpenAI, Anthropic and others have not all signed onto Hassabis’s exact FINRA-style structure. His argument is that the world has only a narrow window to set guardrails before AGI arrives in a few years, but the real fight is over who gets to write those guardrails first.