Why DeepMind's CEO is Calling for US-Led Frontier AI Tests

Demis Hassabis, CEO of Google DeepMind, has urged the establishment of a regulatory body under US leadership to assess frontier-class AI models.

The Nobel laureate warned that immediate measures are necessary to manage risks spanning nuclear and biological threats as the industry moves closer to artificial general intelligence (AGI).

Demis co-founded Google DeepMind and received the Nobel Prize in Chemistry alongside colleague John Jumper, who has since moved to Anthropic. He holds a PhD in Cognitive Neuroscience from University College London.

While issuing warnings about AI threats, Demis remained positive about the technology's promise. "We could even reach a point where resources are no longer the limiting factor for human progress, leading to an amazing new era of abundance," he says, noting AI could address major challenges facing humanity.

AGI Competition and Emerging Threats

The crux of the problem, Demis points out, is that the leading AI companies are currently engaged in an intense commercial and geopolitical competition. The race includes Google DeepMind itself alongside competitors OpenAI, Anthropic and DeepSeek.

"We've already seen the challenges frontier models pose for cybersecurity and other threats including nuclear and bio risks may soon emerge as capabilities continue to advance," Demis wrote on his Substack blog. 

"We must use this precious window before AGI arrives to shape this technology for the benefit of all humanity."

Demis noted that AI threats could emerge from areas currently unknown. "On the horizon, we will need robust safeguards to maintain control of increasingly agentic, recursively self-improving systems – and tackle unknown issues that will only become clearer over time," he stated.

His comments follow the US Government's national security directive that temporarily suspended foreign national access to Anthropic's Fable 5 and Mythos 5 models. The restrictions have reportedly been lifted.

The Center for International Relations and Sustainable Development published an article titled Why AGI Should be the World's Top Priority. It warned that without national and international regulation, humanity could lose control of non-biological intelligence beyond human comprehension.

The Bulletin of the Atomic Scientists set the Doomsday Clock to 85 seconds to midnight in January 2026. This marks the closest position to midnight in the clock's history and includes calls for urgent international AI guidelines.

The proposed US-led standards body

Demis argued that the US holds a strong position to develop a framework for testing frontier models, given its economic and technical capabilities. His proposal calls for a dynamic and rigorous approach to model assessment.

"It could establish a new Standards Body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority," Demis wrote. 

"The Standards Body would be responsible for developing assessment protocols and working with appropriate federal agencies and the US National Labs to conduct testing in areas relevant to national security.

"Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release. Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market."

Demis stated that a US-initiated effort would provide a foundation for shared international standards on frontier AI. Nik Kairinos, CEO and Co-founder of RAIDS AI, responded to the development with caution.

"A global AI watchdog cannot be led by the agenda of any one nation," Kairinos says.

"AI does not respect borders, and nor should the rules designed to govern it. A multinational approach is essential to ensure any AI regulations achieve the necessary trust, cooperation and buy-in to work internationally."

Alignment failures in frontier AI models

Demis's warning comes after Anthropic published a report titled Agentic Misalignment in Summer 2026. The report details observations of alignment model failure in high stakes simulations using frontier models from Google DeepMind, OpenAI, DeepSeek and Anthropic.

The misalignment issues include models pursuing their own motivation beyond user instructions and what Anthropic terms 'harmful compliance'. This occurs when the user's request is followed but the request itself is harmful.

Another issue Anthropic identified is covert sabotage – an agentic misalignment failure where models interfere with code to undermine user intent. Some models also assisted users with conduct that appeared to be white-collar crime.

Additional issues from frontier models included coaching human proxies to whistleblow, where models leak confidential safety information externally or steer humans toward doing it for them. Motivated mislabelling was also observed in the simulations.

Anthropic wrote on its Alignment Science Blog that the purpose of its case studies was to enable developers and evaluators to measure similar failures. The company aims to support the development of targeted safeguards against these alignment issues.