[Bug] Fable 5 dual-use safeguard false-positives on legitimate defensive-security (WAF) development, forcing repeated Fable 5 -> Opus 4.8 switches (5 days)

Summary

Claude Code's Fable 5 model repeatedly trips the dual-use safeguard and auto-switches to Opus 4.8 during ordinary defensive-security development. The flagged content is legitimate WAF / detection-engine code (rate limiting, IP banning, attack-pattern regexes used as a detector). This has happened consistently for 5 days and makes Fable 5 effectively unusable for this project.

Environment

• Claude Code: 2.1.233
• Model: Fable 5 (claude-fable-5), auto-switching to Opus 4.8 (claude-opus-4-8) on flag
• Project type: an open-source WAF / security-middleware engine library (purely defensive security)
• Platform: macOS (Darwin 25.6.0)

What happens

Mid-session, on reading or discussing normal security-library code, Claude Code prints:

Fable 5's safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Switched to Opus 4.8.

The switch fires with no dual-use content present. A concrete trigger from today: a message discussing an auto-ban configuration finding in a rate limiter (that _handle_rate_limit_exceeded never calls ban_ip, so auto_ban_threshold / auto_ban_duration are effectively dead config for a rate-limit-only consumer). That is a code-review observation about a rate limiter, not dual-use content.

The safeguard vocabulary of this project is inherently security-flavored: detection patterns, injection payloads carried as a test corpus, rate limiting, IP allow/deny lists. The safeguard appears to fire on that domain vocabulary itself rather than on any harmful intent or capability.

Impact

• Fable 5 is effectively unusable for defensive-security work: the safeguard fires on the domain itself.
• The forced model switch changes model capability mid-task, silently, which is disruptive for multi-step agentic work and breaks continuity.
• 5 consecutive days of repeated flags on the same defensive-security project.
• Reporter is a partner-program (CVP) participant; the ongoing friction is significant.

Expected

Legitimate defensive-security and code-review tasks should not trip the dual-use safeguard. Requests:

1. Materially reduce the false-positive rate on clearly-defensive security code (WAF, rate limiting, detection engines, IP management).
2. Make the model switch recoverable and visible rather than silent, and ideally opt-in, so a false positive does not silently downgrade capability mid-task. (/config exposes a model-switch behavior toggle, but the default still fires on legitimate work.)
3. Provide a per-project or per-session allowance for verified defensive-security repositories.

Frequency

Consistent, multiple times per day, for 5 days, on the same project.

Related issues

This appears to be a widespread, actively-reported problem. Filing a distinct report for the defensive-security-tooling angle (a WAF engine repo where the whole domain vocabulary trips the flag) plus the 5-day duration and partner-program context, but linking the cluster for consolidation:

• #85414 Fable 5 safeguards flag defensive security audit output (false positive)
• #86539 Fable 5 Model Safeguards Overly Flagging Legitimate Coding Tasks
• #83641 Fable 5 safeguards blocking legitimate code operations
• #85303 Unintended model switch to Opus due to Fable safeguard flag during conversation
• #86095 [Feature Request] Add safeguard override or exception handling for legitimate coding tasks
• #84282 verbatim safeguard-message title