pragma.vision Technology observatory

Verification register Security & Identity

Living definition

What is DeepMind Frontier Safety Framework v3 (Tracked Capability Levels)?

AI Safety, Eval & Alignment Security & Identity

As of
2026-07-23
Revision
2026-07-26.0
Method
v1.3.0

Definition

The term, in context

AI-assisted draft · approved dataset

The Frontier Safety Framework is DeepMind's internal protocol for identifying capability thresholds at which an advanced model could enable severe harm, and for binding specified mitigations to each threshold before it is crossed. Its third version organizes those thresholds as capability levels spanning areas such as cyber operations, biosecurity uplift, and a model's potential to undermine human oversight, with each level re-assessed as models are trained and evaluated.

Live readiness status

Status as of the current dateline

AI-assisted assembly · derived results

As of 2026-07-23, verified readiness is 61 (claimed 75, reported 69, gap 14) — Strong evidence strength; current signals suggest Track; not yet.

The readiness fact belongs to the canonical Readiness Verdict for DeepMind Frontier Safety Framework v3 (Tracked Capability Levels).

Dataset approval

Human editorial release

pragma.vision editorial — standing authorization (operator dev@soft.house, 2026-07-17)

Dataset
2026-07-26.0
Hash
sha256:e2eb667df093db8819d579dff9aad51021b4977da85db2dfa82f63f9ed4e72b7
Approved
2026-07-26

Your opinion

Tell us anything.

What works, what doesn't, what's missing — especially about our watches, lenses, and the register itself. Anonymous is fine; leave an email if you'd like a reply.