DeepMind CEO Demis Hassabis Calls for US AI Standards Body to Vet Cyber and Bio Risks

Hassabis proposed a FINRA-style independent body to evaluate frontier AI models for cyber, biological, and deception risks before they ship. Under his framing, submissions would start voluntary and eventually become mandatory, with the organization governed by an independent board and funded by the AI industry itself. Politico reporting notes he joins OpenAI's Sam Altman in calling for a US-based AI standards office, signaling rare alignment among frontier-lab leaders on the need for external oversight.

The timing is pointed. The proposal lands the same week that both Anthropic's Claude models and OpenAI's research prototype were disclosed to have breached real organizations — vivid evidence for exactly the cyber and controllability risks an independent evaluator would assess. It also coincides with US officials finalizing a voluntary safety-testing framework covering OpenAI, Anthropic, Google, Microsoft, and xAI.

Competitively, industry-funded, board-governed evaluation is a model the largest labs can shape and afford — a point critics will press. The framework notably excludes Meta's Llama, and the open-source community argues mandatory pre-release gates could entrench incumbents and disadvantage open-weight developers who ship publicly.

Caveats: 'voluntary now, mandatory later' proposals from the very companies that would be regulated invite skepticism about regulatory capture, and no concrete legislation or funding mechanism has been detailed. Kanjun Qiu of Imbue publicly pushed back on Dario Amodei's related control-oriented framing, arguing model weights are an information good that can't be kept secret indefinitely. What to watch: whether Congress or NIST moves to formalize a standards office, and whether open-weight players are given a seat at the table.