If you discover a security vulnerability in cc-guard, please report it responsibly:

1. Do NOT open a public issue
2. Use GitHub's private vulnerability reporting
3. Or email: goosull@users.noreply.github.com

I'll respond within 48 hours and work with you on a fix.

cc-guard is a permission guard, so its own security matters:

• Fail-open: If cc-guard crashes or can't read rules, it exits 0 (allow). A broken guard never blocks your work.
• Regex only at runtime: No network calls, no AI/LLM calls, no external dependencies during check. Pure regex matching.
• Deny beats allow: A command matching any deny rule is blocked regardless of allow rules.
• Input normalization: Whitespace trimming, newline stripping, and compound command splitting reduce evasion vectors.
• LLM calls only in learn: The learning engine calls an LLM API, but only when you explicitly run cc-guard learn. Never during hook execution.

• Regex patterns don't prevent all evasion (e.g., shell variable expansion: $CMD where CMD=rm -rf /)
• Subshell expressions $() and backticks are not yet decomposed
• Unicode homoglyph attacks are not normalized