[Bug] Opus 5 cyber safeguard incorrectly blocks defensive security review of user's own code

Bug Description
Opus 5 cyber safeguard blocks defensive security review of my own code

What happened:
In Claude Code, I asked Opus 5 to run a one-round security review (the
security-review skill) of a local diff to my own private skill: a reader
that turns my own EPUB/PDF books into notes. The session was paused
mid-review with "Opus 5's safeguards flagged this message", Details: [cyber].
No findings were written. The only options were "Switch to Opus 4.8" or
"Edit prompt and retry with Opus 5".

What the review asked for:
Check that my parser fails safely on malformed files: zip size limits,
XML entity handling in EPUB metadata, deeply nested PDF bookmarks, hidden
Unicode surviving text extraction, and clean errors instead of tracebacks.
Sample files under /tmp only, no edits, no network, no third-party targets.

Why this is a problem:
This is standard defensive review of code I wrote and run on my own
machine. Security review of your own parsers is exactly what a coding
agent should help with. Blocking it pushes people to skip security review
or downgrade models, which makes their code less safe, not more.
Hitting "continue" did not help; the session stayed paused.

Expected:
Opus 5 finishes a defensive review of the user's own code. If a
classifier is unsure, it should ask for context, not hard-pause the session.

Environment: Claude Code CLI, Opus 5, macOS, 2026-09-16.

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.273
• Feedback ID: a2433a4a-7ee6-41a1-ad73-422f328344d0

Errors

[]