Testing CBRN and Hazardous Capability Safety

What to do

1. Evaluate whether the system increases harmful capability, refuses or redirects appropriately, avoids operational detail, becomes more dangerous with tools or retrieval, and still supports benign education.
2. Build paired cases that differ in intent, requested specificity, authorization, and context.
3. Score helpful safety guidance, harmful capability uplift, unsupported assumptions, escalation, and consistency across paraphrases.
4. Measure capability uplift, operational specificity, refusal quality, benign over-refusal, tool amplification, retrieval amplification, and post-release drift.
5. Use benchmark families such as WMDP, AILuminate, HarmBench, JailbreakBench, CyberSecEval, CyberSOCEval, METR autonomy evals, and scheming evals as anchors, not proof of safety.

Evidence to preserve

• Preserve the inputs, versions, configurations, raw outcomes, and results for hazardous capability, CBRN, cbrn hazardous capability safety needed to reproduce work on Testing CBRN and Hazardous Capability Safety.
• Report results for hazardous capability, CBRN, cbrn hazardous capability safety by relevant slice, separate blocker failures from averages, state uncertainty and blind spots, and connect the result to a release decision.