DeepMind CEO calls for a US-led group to test frontier AI models

In short: DeepMind chief executive Demis Hassabis wants a US-led group to test the most advanced AI models for security and safety risks.

What happened

Demis Hassabis, who runs Google DeepMind, published an essay calling for a new standards body in the United States. He said the goal should be to test “frontier” AI models, meaning the most capable new systems, for threats such as cyber attacks and other national security risks.

He argued that AI is improving quickly and that regulators need “urgent action.” He also said society has a “precious window” to prepare. Hassabis wrote that artificial general intelligence, which he described as AI that can match all the capabilities of the human brain, could be only a few years away.

Hassabis suggested the new body could be modeled on FINRA, a US self-regulatory organization that oversees parts of the financial industry to protect investors. In his proposal, the body would review frontier AI models made anywhere, including “open” models that can be downloaded and “closed” models that are controlled by a company.

He also said the group should “stress-test” safety guardrails, meaning it should try to find ways to get around built-in safety rules, like checking whether a lock is easy to pick (testing the lock by trying to break it). The Financial Times reports that concerns about bypassing guardrails helped prompt a recent US export ban affecting Anthropic’s most advanced models.

Why it matters

AI models are starting to be treated more like critical infrastructure, not just consumer software. If governments and companies agree on shared tests, it could shape which AI systems are allowed to be sold or shared across borders, and how safe those systems need to be before the public uses them.

Source: Financial Times