[Bug] False-positive report

Bug Description
false-positive report

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.172
• Feedback ID: 892819f9-40f9-456c-b20d-9b7abb511ecf

Errors

⏺ Fable 5's safety measures flagged this message for cybersecurity or biology topics. They may flag safe, normal content
  as well. These measures let us bring you Mythos-level capability in other areas sooner, and we're working to refine
  them. Switched to Opus 4.8. Send feedback with /feedback or learn more
  ⎿  Tip: You can configure model switch behavior in /config

[Bug/false-positive report]

Environment: Claude Code, model Claude Fable 5 (claude-fable-5[1m], 1M context), Japanese-language session.

What happened: After several turns discussing software security-architecture topics (approval gates, audit trails, tamper-evidence, threat modeling for our own product), the Fable 5 safety measure fired and auto-switched to Opus 4.8 — but it fired on a harmless one-line user message ("Difyみたいな感じ？" / "Is it like Dify?") that contained NO cybersecurity or biology content.

Why this looks like a false positive / possible bug, not expected behavior:

1. The trigger turn's input text contains no security/biology terms.
2. The security-heavy discussion was several turns earlier; the flag fired on the least sensitive turn.
3. Suggests the classifier reacted to accumulated conversation context or internal state rather than the current message.

Impact: The discussion was entirely legitimate product/business design (building a governance layer for AI agents). The model switch was seamless but unexpected.

Request: Please review whether the cybersecurity/biology safety measure is over-triggering on legitimate security-engineering discussions, and whether the trigger is evaluating the current message vs. accumulated context. Happy to provide the session if useful.