[Bug] Model switching due to safety policy blocking security mitigation discussions

Bug Description

Opus 5.5 and Fable safeguards repeatedly flag a defensive security-remediation task as a cybersecurity risk, and the session switches to a fallback model (Opus 4.8).

Context

• Task: writing a remediation report and internal presentation for an LLM-based NL2SQL service before its pilot launch (internal, closed-network deployment).
• Internal QA testing on our own dev environment is already complete. No further vulnerability discovery or testing is being requested.
• Goal of the flagged prompts: research public defensive guidance (OWASP Top 10 for LLM Applications, guardrail documentation, Text-to-SQL access-control best practices) and draft remediation plans and a presentation.

Steps to Reproduce

1. Start a session with Opus 5.5 (also reproduced with Fable).
2. Ask for a research plan and remediation report on securing an internal LLM NL2SQL service, framed explicitly as defensive ("findings are already collected; only remediation is needed").
3. The session is flagged and switched to Opus 4.8.

Rewording the prompt (removing words like "vulnerability", focusing on "guardrails" and "improvement plan") and attaching only categorized results without raw test inputs did not help.

Expected Behavior

Defensive remediation work on one's own service (guardrail design, access-control improvements, reporting) should be allowed on Opus 5.5 / Fable, or the flag should explain what triggered it so the user can adjust.

Actual Behavior

Repeated flags with no indication of which part of the context triggered them. It is unclear what to change.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.290
• Feedback ID: 67ce0bc5-2245-4792-b375-7b83a0de6b75