The AI Safety Framework
Four risk levels from the published policy, with increasing evaluation, approval, and review requirements. Classification depends on assessment; the selected level is an example.
High Risk
Restrict access and require external challenge and governance review for high-risk systems.
Applicability
- Systems assessed as high risk
- CBRN-relevant capabilities require an applicable expert gate
Measures
- External red teaming
- Restricted access
- Governance Review Board approval
- CBRN expert gate where applicable
- Quarterly and post-incident review
"A system assessed as high risk would remain access-restricted while external red teaming and Governance Review Board review are completed. This example does not establish a current product classification."
Core Principles
The eight foundational pillars that guide responsible AI deployment across all capability levels.
Capability‑Risk Alignment
Safety requirements scale with model capability.
Clear Thresholds
Measurable triggers signal transitions to the next safety tier.
Layered Safeguards
Combine technical, procedural, and organizational measures.
Rigorous Evaluation
Formalize testing, including adversarial probes and red teaming.
Accountable Governance
Establish oversight bodies to review and approve deployments.
Continuous Monitoring
Deploy real‑time telemetry to track model behavior.
Transparent Reporting
Document assessments, incidents, and mitigation steps.
Research Integration
Stay current with AI safety research and incorporate it.
Implementation Roadmap
A structured, step-by-step approach to integrating the safety framework into your organization.
Capability Assessment
Catalog your AI projects and map each to the corresponding safety level.
Gap Analysis
List required safety measures that are not yet implemented and prioritize them by risk.
Roadmap Development
Create a timeline for implementing missing safeguards, starting with the highest-risk projects.
Governance Structure
Establish safety review committees with clear roles, responsibilities, and decision-rights.
Measurement & Monitoring
Build dashboards to track key safety metrics and set up alerting thresholds.
Training & Culture
Educate all teams on the framework and foster a culture of "safety by design."
Iterate & Evolve
Review framework thresholds and measures quarterly, incorporating new safety research and lessons learned.