System cards/google-deepmind

Google DeepMind

Gemini family of models from Google DeepMind. Frontier Safety Framework tracks Critical Capability Levels across five domains: autonomy, biosecurity, cybersecurity, ML R&D, and deception.

Canonical site ↗Seoul signatoryParis signatory
Framework versions (Frontier Safety Framework)
  1. 2025-02-04FSF v2.0pending

    v2 expanded the FSF to a fifth domain (deception) and tightened CCL definitions. Pending verification of exact text against the published v2 document.

    • Added deception as a fifth tracked CCL domain
    • Tightened biosecurity CCL definitions
    • Pending verification of the full change set
    Primary source ↗PDF ↗
  2. 2024-05-17FSF v1.0pending

    Initial Frontier Safety Framework. Established Critical Capability Levels (CCL) across autonomy, biosecurity, cybersecurity, and ML R&D domains.

    • First public Frontier Safety Framework
    • Defined Critical Capability Levels (CCL) per domain
    • Set evaluation cadence and mitigation requirements
    Primary source ↗PDF ↗
Model cards
  • 2025-11-19against FSF v2.0pending
    Gemini 3 Pro
    Capability: CCL-1 across all five domains
    Vertex AI + Gemini app + AI Studio
    Autonomyccl-1Biosecurityccl-1Cybersecurityccl-1ML R&Dccl-1Deceptionccl-1
Other labs tracked