Demis Hassabis calls for U.S.-led AI standards body to test frontier models

Google DeepMind chief Demis Hassabis called on Tuesday for the U.S. to establish a federally overseen public-private partnership to set standards and test frontier AI models for national security risks before public release. The proposed body would function similarly to FINRA (the financial regulator) and include independent technical experts and open-source representatives on its board. Hassabis argued that frontier labs would initially voluntarily submit models for 30-day review before deployment, with the framework becoming mandatory after demonstration of effectiveness. Specific testing would include attempts to bypass safety guardrails, signs of deception, and assessment of practices like digital watermarking and human-readable model outputs.

Hassabis acknowledged that the standards body would require 'substantial' funding to attract world-class talent and computational resources, with costs likely shouldered by the industry. The proposal directly references threats including cybersecurity risks and potential nuclear and biological hazards as AI capabilities advance toward AGI. This call mirrors similar proposals from Anthropic CEO Dario Amodei and OpenAI's Sam Altman, both raised at a June G7 meeting with President Trump and global leaders.

For architects, Hassabis's framing of a U.S.-led oversight framework carries immediate policy weight: it positions the U.S. as the regulatory standard-setter in an era when Chinese models (DeepSeek, Z.ai) are gaining traction with American companies precisely because of cost and API flexibility. A federated testing body would likely become the de facto gating mechanism for high-capability model releases in the U.S. market. The open questions remain funding, governance structure, and whether frontier labs will voluntarily comply or face export restrictions if they don't.