Fable 5 [cyber] safeguard false-positives make the model unusable for telecom/VoIP development

Status Fixed / completed
Maintainer reply None cached
Activity 0 comments · opened Aug 17, 2026 · closed Aug 17, 2026

Summary

Fable 5 is effectively unusable for coding on my project: its safety classifier repeatedly flags legitimate software-engineering sessions with Details: [cyber] and force-switches the session to Opus 4.8 mid-task. This has happened consistently — I have never managed to complete a coding task on Fable 5 in this codebase.

Context

I work at a Brazilian telecom operator (MVNO). The codebase is ordinary telecom backend work: SIP trunk configuration, voice/call APIs, DID number provisioning, WhatsApp Business onboarding, webhook handling. Nothing offensive-security related.

The most recent flag happened in the middle of building a large feature (a voice/call API, "Fase 0: SIP tests"). The session involved normal domain vocabulary — SIP, UDP, API tokens, credential configuration, sandbox network restrictions — and the safeguard fired mid-conversation:

"As proteções de Fable 5 sinalizaram esta mensagem... Alternado para Opus 4.8. Detalhes: [cyber]"

The problem

  1. Neither I nor Claude itself can tell what triggered it. The only detail exposed is the string [cyber]. When I asked Claude (Fable 5, in a new session) why the previous thread was flagged, the best it could do was informed guessing: telephony vocabulary (SIP/UDP/tokens/credentials) probably pattern-matches credential-harvesting or network-tooling classifiers.
  2. It's a hard mid-task interruption. The model switch lands in the middle of multi-step orchestration (worktrees, task lists, spawned subagents), which is exactly when you least want the model swapped underneath you.
  3. For a telecom domain this is not an edge case. Every meaningful session in this codebase touches SIP, tokens, and network configuration. If that reliably trips the classifier, Fable 5 is simply not usable for an entire industry vertical of legitimate work.

What would help

  • Expose what in the message tripped the classifier (even a rough category or the offending span), so users can rephrase or restructure.
  • An in-product way to say "this is a false positive, keep me on Fable" — an appeal/override for authenticated users on their own private repos.
  • Better calibration for telecom/VoIP vocabulary: SIP + UDP + token/credential configuration is the daily bread of a large legitimate industry.

Environment

  • Claude Code desktop app on Windows 11
  • Model: Fable 5 (claude-fable-5), auto-switched to Opus 4.8
  • Flag detail shown: [cyber]

View original on GitHub ↗