[MODEL] Fable 5 safeguards false-positive: routine SSH administration of own infrastructure consistently reroutes sessions to Opus

Summary

Fable 5's dual-use safeguards consistently flag routine administration of our own infrastructure over SSH as offensive cybersecurity activity and reroute the session to Opus. This makes Fable 5 effectively unusable for ordinary DevOps/sysadmin work, which is most of what we use Claude Code for.

Preflight

• Searched existing issues. Closest reports: #74783 (git/diagram work flagged) and #73784 (anti-fraud T&S work flagged). This report is about plain first-party infrastructure administration (SSH + Docker + logs), which seems to be a distinct and very common false-positive class.
• This report contains no sensitive information.

Type of behavior issue

Safeguard false positive → forced model fallback (closest dropdown option: "Claude refused a reasonable request").

What we ask Claude to do (typical, benign)

• SSH into our own servers (a Synology NAS and a few small on-prem hosts we own) to check status: docker ps, read container logs, run health-check curls against our own APIs.
• Restart our own Docker containers after editing our own application code (Python/Bottle, FastAPI, Next.js).
• Read-only SQL queries against our own ERP databases.
• Ordinary editing and deployment of our in-house apps.

No exploit development, no malware, no scanning, no access to third-party systems. It is the daily operation of a small company's self-hosted infrastructure, performed by its owner.

What happens

In most working sessions on this infrastructure, the safeguard banner appears (Spanish UI):

"Las salvaguardas son intencionalmente amplias en este momento y pueden marcar trabajos seguros y habituales de programación, ciberseguridad o biología. Estas medidas nos permiten ofrecerte las capacidades de nivel Mythos antes, y estamos trabajando para perfeccionarlas. Se ha cambiado a Opus 5."

and the session is switched away from Fable 5. (Side note: the banner says "Opus 5" while support article 15363606 says the fallback is currently Opus 4.8 — one of the two seems out of date.)

Our hypothesis: the classifier reacts to the shape of the context (SSH commands, hostnames, key paths, Docker operations, log output accumulating in context) rather than intent, and treats infrastructure administration as offensive cyber tooling. Since the checks review everything in context, a session that starts fine degrades as soon as normal ops content accumulates.

Expected behavior

Fable 5 continues the session. Operating your own servers over SSH is systems administration, not "offensive cybersecurity techniques".

What we already tried

• Completed the Cyber Verification Program application (portal.anthropic.com/programs/cvp). Per support article 14604842, CVP applies to the real-time cyber safeguards on Opus/Sonnet only — it appears to have no effect on Fable 5's routing, which was a surprise to us as a verified user.
• Using the in-product "Send feedback" option on the fallback notice when it appears.

Request

1. Tune the Fable 5 cyber classifier to distinguish administration of one's own infrastructure (SSH to known hosts, Docker ops, log reading, health checks) from offensive tooling; or
2. Honor CVP verification (or an equivalent declared-use-case mechanism) in Fable 5's routing decision, so verified users doing legitimate operations work can keep Fable 5.

Environment

• Claude Code CLI 2.1.211, running both interactively and headless (claude remote-control) in a Docker container on our NAS; same behavior from the Claude desktop/mobile apps.
• Model: claude-fable-5 (session default). Max plan (claude.ai OAuth).
• Spanish-language UI.