False-positive safeguard flag interrupted legitimate work; /feedback unavailable in Cursor harness

Safeguard false-positive report — Claude Code (Cursor harness)

Environment: Claude Code running inside Cursor (not standalone CLI)
Model at time of flag: Fable 5 (claude-fable-5), auto-switched to Opus 4.8
Date: 2026-07-15
Note: /feedback is unavailable in the Cursor harness, hence filing manually.

---

False positive. I'm the founder of a software consultancy working in my own private repositories and my own business strategy throughout this session. The flagged work was: reviewing my own codebases, defining IP boundaries between my product and client deliverables, and defensive credential hygiene — rotating and scrubbing API tokens that had been accidentally committed to a private repo I own. This is routine authorized security remediation on my own systems, not anything adversarial.

The classifier interrupted a live working session and switched models mid-task. I understand the safeguards are deliberately broad right now, but flagging a business owner discussing and securing his own software is exactly the kind of false positive that makes the tool feel unusable for real work. Please tune for authorized, self-owned defensive-security and business context.