Google / Wikimedia Commons (Public domain)
Demis Hassabis backs a slower pace for frontier AI as safety researchers quit and warn of catastrophic risks
The person who helped build some of the most powerful AI systems on Earth is now saying the industry needs to pump the brakes. Demis Hassabis, co-founder and CEO of Google DeepMind, publicly endorsed slowing down frontier AI development in September 2026 to give safety research a chance to catch up with rapidly advancing capabilities.
His comments align with a growing chorus of alarm from inside the labs themselves, where researchers are not just raising concerns but walking out the door.
The case for slowing down #
Hassabis’s position echoes an essay from Anthropic CEO Dario Amodei, who argued that AI capability gains are at risk of outrunning the control mechanisms meant to keep them in check.
Back in July 2026, Hassabis proposed the creation of a US-led, industry-funded standards body designed to conduct pre-release testing of frontier models. The focus areas: cybersecurity vulnerabilities, biological threat potential, and agentic behavior, meaning AI systems that can act autonomously in the world with minimal human supervision.
Safety researchers are voting with their feet #
Bilal Chughtai, a former AGI safety researcher at DeepMind, resigned in July 2026. By mid-September, he was publicly warning that AI has the potential to “kill us all” and that the window for developing adequate alignment methods is closing fast.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Josh Engels, another member of DeepMind’s AGI safety team, left in early September 2026 to join METR, an organization focused on evaluating AI risks. Engels cited what he called a “terrifying chance” of catastrophic outcomes from self-improving AI systems, and suggested those outcomes could materialize within five years.
An open letter signed by over 1,300 researchers from frontier AI labs has highlighted the same fundamental tension: capabilities are advancing faster than the field’s understanding of how to control them.
What a standards body might actually look like #
The three focus areas Hassabis identified, cybersecurity, biological risks, and agentic behavior, represent some of the highest-consequence failure modes. An AI system that can autonomously discover novel cyberattack vectors or assist in designing dangerous biological agents is a qualitatively different kind of risk than one that generates a bad essay or hallucinates a legal citation.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our