[Bug] Cyber safety classifier false positives on config review and flag discussion
Bug Description
Repeated false positives [cyber] while discussing the flag itself (Opus 5.5)
Context: same company and product as my other report (AutomaDesk, our own fork of
RustDesk for customer support).
What happened:
- I asked Claude to review my global CLAUDE.md on my own 4 machines (Mac, Linux
server, Windows PC, a colleague's Mac) over our own SSH aliases. The goal was to
check that the work-context section is consistent. Claude read and compared the
files. Its response was then stopped by the safety classifier.
- I then asked why the previous session had been fully blocked when this one was
allowed to continue. That answer was stopped too.
- Even an answer that only summarized the public support article
(support.claude.com/en/articles/8106465) was stopped.
Being unable to even ask about a flag makes it hard to understand it or work around
it. The tasks involved were reading config files on my own machines and asking about
the safeguard behavior. None of it was offensive.
Environment Info
- Platform: darwin
- Terminal: Apple_Terminal
- Version: 2.1.294
- Feedback ID: e19c5150-4db1-4475-89e2-d168bc6d9b67