Hassabis warns bio, nuclear risks in open models within 18 months

Demis Hassabis picked a specific timeline this week, and the specific part is what ought to focus minds. The DeepMind co-founder and CEO told Axios that today's AI-driven cyber risks are 'warning shots,' and that within 18 months those capabilities, plus far graver biological and nuclear threats, could live inside open-source models beyond any government's control. He laid it out in a personal manifesto published Tuesday morning, 'A Framework for Frontier AI and the Dawning of a New Age.'

The proposal itself is unusually specific for a CEO op-ed. Hassabis wants an AI standards body modeled on FINRA, the private, industry-funded watchdog that polices Wall Street under SEC oversight, funded by industry, staffed by world-class technical experts, and answerable to the US government. Frontier labs would voluntarily share their models with the body up to 30 days before release, for safety testing that probes dangerous cyber, biological and 'deception' capabilities. He wants it stood up by year's end.

For readers outside a frontier lab, the bet is that a voluntary, US-anchored pre-release screen can front-run open-weight distribution of the same capabilities. If the 18-month claim is directionally right, the window to build that screen closes faster than any legislative process usually moves. As TechTimes reported, the body would also carry the power to coordinate an industry-wide slowdown if evaluations warrant it, which is a sharper tool than anything in the current safety-pledge landscape.

Read it as one CEO's manifesto paired with an exclusive interview, not a signed accord. Take the 18-month timeline as Hassabis's estimate, not settled forecast, and take the FINRA analogy as a proposal, not a mandate. The reporting does not include participation commitments from OpenAI, Anthropic, Meta, xAI or Chinese labs, coordination with the White House, or how any of this would touch open-weight releases already in the wild.

Even so, a Nobel-laureate CEO publicly asking for a screen on his own industry, on a countdown, is the sort of signal that tends to pull policy conversations forward faster than the industry expects. It lands the same day House Democrats grilled OpenAI and Anthropic on rogue AI agents, one of 218 AI safety stories we have logged in the last 90 days.