Thirty days of review before a frontier model is released

Demis Hassabis suggests testing frontier models for thirty days. A mandatory evaluation to secure artificial intelligence systems.

Demis Hassabis, head of Google DeepMind, is proposing a regulation framework for frontier AI: the creation of a US standards body tasked with evaluating the most powerful models before they are released to the market.

The envisioned setup is inspired by FINRA, the self-regulatory authority for US financial markets: a public-private partnership under federal supervision, with a board bringing together independent technical experts and open-source representatives, and funded primarily by the industry in order to attract top-tier engineers and pay for the computing power required for large-scale testing.

The mechanism relies on a threshold. A model would be deemed "frontier" if it surpasses certain benchmarks on a battery of evaluations defined and regularly updated by the organization. The companies involved would become "frontier labs," required to adhere to specific practices: publishing technical specification sheets, internal cybersecurity, vetting of key personnel, and dedicated resources for safety research. Initially, these labs would voluntarily submit their models for review up to thirty days before their release; once the protocol is proven, this process would become mandatory for any launch on the US market.

The evaluations would focus on cybersecurity, biological threats, and other high-risk areas, with specific tests for agentic systems: attempts to bypass safeguards and signs of deception. Hassabis also proposes that the framework could, if the situation demands it, coordinate a slowdown in development. The framework would apply regardless of the model's country of origin, whether open or closed, while ordinary models from startups or academia would be exempt.

His justification rests on a single observation: the commercial and geopolitical race is advancing capabilities faster than our understanding of them, and no one, even among experts, agrees on what comes next.