Why it matters

As frontier models get embedded into products, “trust us” safety claims aren’t enough. Public frameworks like FSF shape what customers, regulators, and partners expect to see: explicit thresholds, documented mitigations, and clear go/no-go governance.

What changed

The most important change is operational: DeepMind adds Tracked Capability Levels to watch for meaningful risks below the “critical” threshold, and it extends its taxonomy to include harmful manipulation as a first-class risk domain alongside areas like cyber and CBRN misuse.

Practical read

The accompanying FSF 3.1 document describes how the company plans to use early-warning evaluations, alert thresholds, and safety-case reviews as part of a broader risk acceptance process. Even if you don’t adopt the exact same structure, it’s a useful blueprint for internal gating and disclosure.

What to watch

For organizations buying or deploying AI systems, this is a reminder to ask for artifacts, not assurances: what evaluations were run, what mitigations were required, and what monitoring exists post-deployment.