Strengthening our Frontier Safety Framework

Notes

First-party DeepMind announcement describing its own safety framework; primary_doc prior (0.90) lightly discounted because it is a self-descriptive governance statement, not a binding artifact.

Evidence from this source (5)

• CDDeepMind's Frontier Safety Framework v3 adds a Critical Capability Level for ha… support
  weight

  0.95

  method: primary_testimony · locator: Section: Addressing the risks of harmful manipulation

  “we're introducing a Critical Capability Level (CCL) focused on harmful manipulation --- specifically, AI models with powerful manipulative capabilities that could be misused to systematically and substantially change beliefs and behaviors in identified high stakes contexts”

• CDFSF v3 extends DeepMind's risk assessment process beyond early-warning capabili… support
  weight

  0.85

  method: primary_testimony · locator: Section: Sharpening our risk assessment process

  “Building on our core early-warning evaluations, we describe how we conduct holistic assessments that include systematic risk identification, comprehensive analyses of model capabilities and explicit determinations of risk acceptability.”

• CDUnder FSF v3, DeepMind extends its pre-launch safety case review process to cov… support
  weight

  0.90

  method: primary_testimony · locator: Section: Adapting our approach to misalignment risks

  “For advanced machine learning research and development CCLs, large-scale internal deployments can also pose risk, so we are now expanding this approach to include such deployments.”

• CDDeepMind's FSF v3 replaces the prior exploratory instrumental-reasoning (decept… support
  weight

  0.90

  method: primary_testimony · locator: Section: Adapting our approach to misalignment risks

  “While our previous version of the Framework included an exploratory approach centered on instrumental reasoning CCLs ... with this update we now provide further protocols for our machine learning research and development CCLs”

• CDIn FSF 3.1 (April 2026 update), DeepMind introduces Tracked Capability Levels (… support
  weight

  0.95

  method: primary_testimony · locator: Section: FSF 3.1: Introducing tracked capability levels

  “As of April 17, 2026, we are adding Tracked Capability Levels (TCLs) in certain domains to our Frontier Safety Framework, introducing a new capability level to help us spot and evaluate potential less extreme risks sooner.”