[Bug] Fable 5 dual-use safeguard over-flagging legitimate biomedical/neuroscience content and not auto-recovering model selection
Bug Description
Subject: Fable 5 dual-use safeguard — repeated false positives on legitimate medical / neuroscience work
What happened: While doing routine work (digital-twin engineering + maintaining a personal medical/neuroscience knowledge wiki), Fable 5's safeguard flagged my message and silently switched me to Opus 4.8. The content was ordinary, defensive medical AI / clinical research — not CBRN or any dual-use capability.
Why this is a false positive: The flagged context involved medical-device development and neuroscience literature (Alzheimer's biomarkers, neuroimaging, published clinical reviews). None of it provides uplift for bio/chem/cyber harm. Example of the kind of benign content that likely pattern-matched: a note summarizing a Lancet Neurology rapid review on COVID-19 olfactory dysfunction as a neurodegeneration risk marker — which necessarily mentions SARS-CoV-2 / ACE2·TMPRSS2 receptor entry. This is standard, published, defensive clinical science, but it appears to trip a broad virology classifier.
Two concrete problems:
- Over-broad triggering — mere biomedical terminology (viral-entry receptors in a neurodegeneration context, medical-device development questions) seems sufficient to flag, with no actual dual-use intent or capability.
- No automatic recovery — after switching to Opus, the session did not return to Fable on subsequent, unrelated messages. I had to manually /model back to Fable each time. Documentation/UX suggested it would return automatically; it does not.
Impact: Frequent interruptions and manual model-switching during legitimate medical/engineering work; erodes trust in the safeguard's precision.
Ask: (a) Tighten the biology classifier so defensive/clinical/diagnostic contexts (medical devices, published neuroscience, disease biomarkers) aren't flagged on terminology alone; (b) clarify or fix the post-switch recovery behavior — either auto-return on clean turns, or document that recovery is manual.
Environment Info
- Platform: darwin
- Terminal: iTerm.app
- Version: 2.1.198
- Feedback ID: 970f3348-7024-4743-8d1a-3b79a13d392b
Errors
[]