[Bug] Fable 5 safety classifier false-positive on defensive security code triggers unwanted Opus 4.8 fallback
Bug Description
Subject: Fable 5 safety classifier false-positive → repeated unwanted fallback to Opus 4.8 on defensive-security software work
We are using Claude Code (Fable 5, multi-agent workflow) to build a French accounting copilot — a normal software engineering
project. The cyber/bio safety classifier repeatedly flags our sessions and silently falls back to Opus 4.8, which breaks our
chosen-model workflow and forces a manual /model fable re-switch each time.
The flagged content is defensive application security on our own codebase, not offensive cyber or bio content. Concretely, the
conversations that trigger the fallback discuss:
- closing authentication/authorization vulnerabilities in our own BFF (JWT signature verification, tenant isolation, fixing an
endpoint that returned data without ownership checks);
- GDPR/PII anonymization of user data before sending it to an LLM;
- adversarial code review of our own code to harden it (we deliberately ask agents to "try to break" our write path, find
double-submission edges, find data-leak paths);
- standard security vocabulary used in a defensive context: "bypass", "injection", "jailbreak", "leak", "sentinel", "fingerprint",
"deanonymize".
None of this is cyber-offensive (no attacking third-party systems), bio, or otherwise harmful — it is a developer hardening their
own product, which is exactly the behavior security guidance encourages. The classifier appears to key on security vocabulary rather
than intent/target.
Request: please tune the classifier to distinguish legitimate defensive security of one's own software (vulnerability remediation,
secure coding, adversarial self-review) from offensive cyber content. The repeated false-positive fallback is disruptive to a
benign, high-value engineering workflow.
Environment Info
- Platform: linux
- Terminal: vte-based
- Version: 2.1.170
- Feedback ID: 834bd435-70da-4da5-87e9-d099355e4a87
Errors
[]This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗