Google DeepMind strengthens the Frontier Safety Framework

Google DeepMind announced on September 22, 2025 that it had published the third version of its Frontier Safety Framework, which it describes as its most comprehensive approach so far to identifying and reducing severe risks from advanced AI models. The company says the update draws on collaboration with experts in industry, academia and government, along with lessons from applying earlier versions. A note says the post was updated on April 17, 2026.

The update contains three main changes. First, it adds a Critical Capability Level for harmful manipulation, covering models with strong persuasive abilities that could be misused to shift people's beliefs and behavior substantially in high-stakes settings, with severe expected harm. Second, it widens the treatment of misalignment, including scenarios in which misaligned models might hinder operators from directing, changing or shutting down their operations. The earlier exploratory approach to deceptive reasoning is supplemented by protocols for capability levels tied to models that could speed up machine learning research and development to destabilizing degrees. Third, it sharpens definitions so that the most serious threats receive the most rigorous governance, and it describes a fuller risk assessment process, including systematic risk identification, analysis of model capabilities and explicit judgments on whether risk is acceptable.

DeepMind explains that when relevant capability levels are reached it reviews safety cases before external launches, and it now extends this to large-scale internal deployments for advanced research and development levels. Mitigations continue to be applied before thresholds are reached.

The April 2026 revision, called FSF 3.1, introduces Tracked Capability Levels in certain domains to detect less extreme risks earlier and gives more detail on risk management from identification to mitigation. The company says the framework will keep evolving with research and stakeholder input.