[Bug] False-positive safety classifier triggers model downgrade during legitimate legal/security work

Bug Description
Subject: False-positive safety downgrade (Fable 5 → Opus 4.8) mid-session

During a single working session the safety measures flagged my session and
auto-switched the model from Fable 5 (Mythos-class) to Opus 4.8. I believe
this was a false positive.

Session context (all legitimate):

• I am preparing a legal defense document, defending myself pro se against a

patent preliminary-injunction petition. The case file unavoidably discusses
the technical subject matter of the dispute — anti-phishing / victim-recovery
software, mobile-app analysis, server interactions, and the relevant Korean
network/communications statutes — purely to argue non-infringement.

• I also had an authorized security-assessment report open in a browser tab as

part of a separate, legitimate engagement.

Likely trigger: accumulation of dual-use security terminology (C2 references,
app decompilation/analysis, network-intrusion law) plus an open
security-assessment/OSINT report — all in a defensive, legal, and authorized
context, not offensive use.

Impact: the mid-task switch was disruptive and arrived with only a vague notice.
None of the activity was offensive security or policy-violating; it was
legal-defense preparation and an authorized assessment.

Request: please review the classifier's sensitivity so that legitimate
legal-defense work and authorized security assessments are less likely to
trigger a Mythos-class downgrade. Happy to provide more detail if helpful.

Environment Info

• Platform: darwin
• Terminal: WarpTerminal
• Version: 2.1.173
• Feedback ID: dd8a52bc-4dbe-4b4a-bb2d-ac0bcf0cf420

Errors

[]