[Bug] Safety classifier stops defensive sandboxing work (seccomp/Landlock) as false positive

Bug Description

# Safety classifier stops — report - Date: 2026-10-14 - Tool: Claude Code (CLI), model claude-opus-5-5 - Session: 62d65fd4-7700-4e78-817b-f4efcd2932de ## What happened In one session, four of Claude's replies were cut off with "stopped by a safety classifier". All four came while working on one feature of my own product: the isolation of the scripts my runtime starts itself. It is a protective feature that limits what those scripts can do on my own machine. 1. While designing that isolation. 2. While looking up identifiers for it in this machine's own system headers (/usr/include). 3. While planning it on top of two open-source libraries instead (pyseccomp and landlock from PyPI, both MIT). 4. While writing this report about the first three. No reason was shown. ## Why I think it is a false positive - The work only restricts processes my own product starts. It does not target any other system or person. - It uses standard, documented Linux kernel features (seccomp and Landlock), which container runtimes, browsers and systemd use for exactly this purpose. - It is just system calls and nothing else. ## Effect The feature could not be built. For now my product runs its scripts without it, which is less safe than what the work would have delivered. ## Request Please review these stops, and allow defensive work on isolating one's own processes.

Environment Info

• Platform: linux
• Terminal: gnome-terminal
• Version: 2.1.296
• Feedback ID: 02cb5636-02f1-4f79-a326-01fb6ab84673