[Bug] Input safeguard over-flags authorized red-teaming work, forcing fallback to Opus model

Bug Description
Fable 5's input safeguard is flagging legitimate, authorized security-competition
work and auto-switching me to Opus 4.8. Context: I'm working on the Kaggle
"AI Agent Security — Multi-Step Tool Attacks" competition, hosted by OpenAI,
Google, and IEEE. It's defensive red-teaming of an open-source, MIT-licensed
benchmark SDK (aicomp-sdk) that simulates a tool-using agent in a local sandbox
with fixture data — no real systems, users, or secrets involved (the "secret" is a
synthetic SECRET_MARKER sentinel), and the goal is to find and document guardrail
weaknesses so agent defenses can be hardened. Winning entries are open-sourced
under MIT.

This is exactly the "safe and routine cybersecurity work" the flag notice says may
be over-flagged. The classifier appears to trigger on the topic regardless of the
benign, clearly-authorized, hosted-competition context, which makes Fable unusable
for this project — every attempt bounces me to Opus. Please narrow the breadth so
authorized red-teaming / benchmark-security research isn't caught.

Environment Info

• Platform: linux
• Terminal: vte-based
• Version: 2.1.207
• Feedback ID: 5b2dc9bd-a278-4bc1-aad1-ac4b55f9c9a5

Errors

[]