The right question to ask about a dangerous capability: not whether a model can explain something, but whether it meaningfully helps someone who could not otherwise have done it. Information freely available in a library provides no uplift. Synthesising, troubleshooting and tailoring it might.
Technical view
Assessed by controlled studies comparing task performance of participants with and without model assistance, against baselines including internet search, measuring time to completion and success rate on proxy tasks. Methodological difficulty is severe: proxy tasks must be safe to run yet predictive of the real thing, and expertise level of participants dominates results. It is nonetheless the framing used in developer safety policies to decide whether a capability threshold has been crossed.
Related
Part of AI Terminology Index, section: Security and Privacy.