[Bug] Claude 3.5 Sonnet safeguards falsely flag legitimate first-party VPN development and code review

Bug Description
Subject: Fable 5 safeguards false-positive on legitimate VPN development and code review

Fable 5 repeatedly flags my fully legitimate, authorized work as a cybersecurity risk and falls back to Opus 4.8.
These are false positives. My context:

1. Developing my own VPN protocol. I'm building a proprietary network/transport protocol (obfuscation, transport,

handshake) for my own commercial VPN product. This is my own IP and product — not circumventing anyone else's systems.

1. Developing my own VPN clients. Cross-platform clients (iOS/macOS/Android/Windows/Flutter) for my own service —

tunneling, network extensions, packet handling. Ordinary product engineering.

1. VPN configuration. Server config, routing, DNS, split-tunneling for my own infrastructure and my own devices.

Administering systems I own.

1. Code review. Security auditing of MY OWN code (auth/JWT, payments, RBAC, provisioning) to find and fix

vulnerabilities in my own platform. This is defensive security on my own systems.

All of this is work on my own commercial product with full authorization. The classifier appears to trigger on
keywords (VPN, protocol, obfuscation, security, vulnerability, auth) without accounting for the fact that this is
first-party product development and defense of my own systems — not an attack on anyone else's.

Request: please calibrate the safeguards so that developing one's own software, administering one's own
infrastructure, and defensively auditing one's own code are not flagged. Right now this materially disrupts my work
and forces constant fallback to Opus 4.8.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.209
• Feedback ID: b31222cf-f27c-480a-93c2-ba869b23e3b0

Errors

[]