[Bug] Anthropic API: Bio safeguard false positives on approved-therapeutic GLP-1 pharmacology content

Bug Description
False positive from Fable 5's bio safeguard. Personal health research on GLP-1 receptor agonists (semaglutide/tirzepatide) — mechanism of action, dosing, and clinical-trial results at the level of patient-facing and review literature — repeatedly triggers the bio flag and forces a fallback to Opus 5. It fires reliably once web-search results or my own prior research notes enter the context, not on the prompt itself, so any multi-turn session on this topic gets downgraded (3 fallback events on 2026-08-04 alone, all apiRefusalCategory "bio"). Example request ID: req_011CdiJXSFC8J2cDLu2zqXTW. GLP-1 pharmacology is mainstream consumer-health territory — millions of patients take these drugs — and nothing in the session approached synthesis or anything dual-use. Please consider tuning the bio classifier to distinguish approved-therapeutic pharmacology from genuinely hazardous content.

Environment Info

• Platform: darwin
• Terminal: xterm-256color
• Version: 2.1.221
• Feedback ID: 18c24f12-a4ca-4019-81c3-8e583c546783

Errors

[]