What
When a model declines a request due to safety measures, the error message is now tailored to the actual reason rather than defaulting to a generic cyber/bio warning for every refusal.
Details
- Refusals specifically triggered by cybersecurity or biology content still display the existing message naming those topic areas.
- All other refusals now show: "This model has measures that flagged something in this session. This sometimes happens with safe, normal conversations." followed by a note that these measures help bring Mythos-level capability in other areas.
- The model-switching notification (shown when Claude auto-falls back to a different model) also uses this same category check, so users who triggered a fallback for non-cyber/bio reasons no longer see an unexplained reference to cybersecurity or biology topics.
- Refusal telemetry now records non-cyber/non-bio categories as
"other"rather than passing through raw category strings, which normalizes internal analytics.
Evidence
New generic refusal string (search for "This model has measures that flagged something in this session"); category classifier introduced (search for H === "cyber" || H === "bio"); updated fallback-switch messages reference the same classifier
Strings lifted out of the shipped bundle, so the claim above can be checked against them.