Hassabis calls for US-led AI watchdog modelled on Wall Street regulator

Google DeepMind chief Demis Hassabis has called for a US-led artificial intelligence watchdog modelled on Wall Street’s financial regulator. In a manifesto published on Tuesday, the Nobel laureate argued that artificial general intelligence is probably only a few years away, presenting a closing window to prepare.

A FINRA for frontier models

The proposed body would operate as a public-private partnership funded by the laboratories themselves. It would be governed by independent experts, including Turing Award winners, alongside government and open-source representatives. Its mandate would be to test the most powerful models up to 30 days before release for cyber-attack capabilities, biological and nuclear risks, and signs of deception.

Participation would initially be voluntary before hardening into a mandatory regime. Any “Frontier-class” model would need to pass these quarterly-updated benchmarks to access the US market, regardless of where it was built or whether it is open-source. Startups and academics would remain exempt.

The power to pause

Hassabis wants the authority to escalate. The body could coordinate a slowdown in development among frontier laboratories if risks are deemed to be growing. He warned that current cyber-risks are merely warning shots, noting that far graver biological and nuclear tools could emerge within 18 months inside open-source models that governments cannot recall.

His timeline is aggressive, aiming to have the body operational before the end of the year. He claims to have briefed the Trump administration, rival labs and European officials for months, receiving “very positive” noises.

Industry pressure

The proposal follows the Trump administration’s sudden export order freezing Anthropic’s most powerful models last month, a move Hassabis described as a wake-up call. OpenAI subsequently held back its GPT-5.6 model until securing government approval. Hassabis and Anthropic chief Dario Amodei have both pushed for a US-led regulatory coalition at the G7, though they differ on whether the regulator should resemble FINRA or the FAA.

Hassabis frames the technology as a discovery on par with fire or electricity. “We’ve essentially found a way to make sand think. It’s miraculous,” he wrote. The unresolved question is whether a body funded by the very labs that profit from these models can truly be trusted to pull the handbrake.