[Bug] Anthropic API: Excessive safeguard flagging on defensive security design discussions

Bug Description
Building GateClear, a self-hosted airport carry-on security-rules service. Ran a multi-agent planning workflow to design a deliberately locked-down Q&A chat feature — the entire goal was to make it abuse-resistant (retrieval-only, no generative answer path, refusal taxonomy). The planning necessarily discussed threat models, prompt-injection, jailbreak vectors and guardrails — standard defensive security design. Right after that workflow, Fable 5's safeguards flagged the session and switched me to Opus 4.8. This was legitimate defensive engineering (designing controls to prevent abuse), not offensive work. The broad dual-use flagging on defensive security-design vocabulary is a false positive worth refining.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.204
• Feedback ID: 476a4143-7797-4962-be46-f0fbb85c2143

Errors

[]