MayGuard

A security auditor for agent skills. Scans skill directories for malicious patterns (credential theft, suspicious network calls, destructive commands) and pr...

MayGuard appears to implement a reasonable local static audit tool, but be aware of its limitations before relying on it: (1) it skips hidden files (dotfiles), so it may miss .env, .ssh/id_rsa, or other credential files — manually check dotfiles or modify the script if you need them scanned; (2) the scoring logic is simplistic (adds 20 points per match) and a comment about per-category caps is inaccurate, so treat the numeric score as a rough heuristic rather than a definitive risk metric; (3) patterns are simple regexes and can produce false positives/negatives, and obfuscated or encoded threats may evade detection; (4) the SKILL.md/strings refer to 'ClawGuard' while the registry name is 'MayGuard' — this mismatch could be a sign of copy-paste or maintenance issues; (5) run the script on extracted skill contents (not on system root) to avoid long scans and to limit scope. If you plan to depend on this tool programmatically, consider reviewing and (optionally) hardening the script: enable scanning of dotfiles, add per-category caps or weighting, and audit the regex list to reduce false positives.

Type: OpenClaw Skill Name: mayguard Version: 1.0.0 This skill bundle, 'mayguard', is designed as a security auditor for other agent skills. The `SKILL.md` clearly describes its purpose and usage, and does not contain any prompt injection attempts. The `scripts/audit.py` code implements static analysis by scanning files for patterns defined in `references/threat_patterns.json`. The script itself does not perform any network calls, execute arbitrary commands, access sensitive host files, or attempt data exfiltration. Its functionality is entirely aligned with its stated benign purpose of detecting malicious patterns in *other* skills, not performing them itself.