[BUG] Fable 5 safety classifier false-positive during benign GPU / container infrastructure debugging (no security or biology request)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

While using Claude Fable 5 (1M) in Claude Code for a routine infrastructure/coding session — debugging AMD GPU (MI50 / ROCm) device numbering inside Podman containers — Fable 5's safety measures flagged one of my messages mid-session and automatically switched the conversation to Opus 4.8. No offensive-security, exploit, malware, or biology request was made; it was ordinary sysadmin/coding work. This happens repeatedly across unrelated, equally benign sessions. The flag is opaque — no indication of which message tripped it or why — so it silently pulls me off the model I'm paying to use.

What Should Happen?

Benign technical / coding / infrastructure work should not trip the safety classifier. With no offensive-security or bio request present, the session should stay on Fable 5 with no downgrade. If a flag does fire, it should state which message triggered it and why, instead of switching silently. The same content is handled by Opus 4.8 and Sonnet 4.6 without any flag.

Error Messages/Logs

Fable 5's safety measures flagged this message. They may also flag normal, safe content. These measures let us bring you Mythos-level capabilities sooner in other areas, and we're working to refine them. Switched to Opus 4.8.

(Original message was shown in French: "Les mesures de sécurité de Fable 5 1M ont signalé ce message. Elles peuvent également signaler du contenu normal et sans danger. Ces mesures nous permettent de vous proposer plus tôt des capacités de niveau Mythos dans d’autres domaines, et nous travaillons à les affiner. Basculé vers Opus 4.8. Envoyer un retour ou en savoir plus.")

Steps to Reproduce

Start a Claude Code session on Fable 5 (1M).
Work on benign infrastructure content: AMD ROCm/HIP GPU enumeration, Podman container device mapping, kernel/PCI device terminology; context files include routine references to credential/secret management.
Continue the multi-turn session.
Mid-session, the conversation is flagged and silently switched to Opus 4.8, with no indication of the offending message.
Note: not deterministically reproducible — the classifier is opaque, which is part of the bug. For trivially reproducible variants, see #66674 (standard academic terms) and the widely reported first-turn "hello" fallback.

Claude Model

Other

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.173 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

It happens with claude-code on ubuntu, claude desktop on windows etc...