Security Monitor

Security Monitor Skill

This skill scans all installed Claude Skills for indicators of compromise and injection attack patterns. It provides automated static analysis to detect malicious patterns before untrusted skills are executed.

What This Skill Does

When invoked, this skill runs scripts/scan_skills.py against all skill directories discoverable from standard install locations:

• ~/.claude/skills/ (personal skills)
• .claude/skills/ (project skills in current working directory)

The scanner checks each SKILL.md and any associated scripts for:

1. External URL references — C2 callbacks, data exfiltration endpoints
2. Credential access patterns — environment variable reads of sensitive keys
3. Persistence indicators — crontab edits, shell RC file writes, skill replication
4. Trigger hijacking patterns — overly broad or authority-impersonating descriptions
5. Suspicious tool grants — unrestricted Bash in allowed-tools without audit trail
6. Hidden content — HTML comments, zero-width characters, base64-encoded blobs

Invocation

To run a full security audit of all installed skills:

Please run the security monitor skill to audit my installed skills.

The skill will execute the scanner and return a structured report. Each skill receives a risk rating: LOW, MEDIUM, HIGH, or CRITICAL.

Scanning a Specific Directory

To scan a skill pack or directory before installing:

Please scan the skills in ./downloaded-skills/ for security issues.

Interpreting Results

• LOW: No significant indicators found. Normal use.
• MEDIUM: Patterns present that warrant review (e.g., external URLs in a skill

that shouldn't need network access).

• HIGH: Multiple indicators or a single high-confidence indicator. Do not execute

until manually reviewed.

• CRITICAL: Strong evidence of malicious intent (credential exfiltration,

persistence, C2 patterns). Quarantine immediately.

Important Notes

• This skill itself requires Bash access to run the Python scanner.
• The scanner is read-only and does not modify any skill files.
• False positives are possible — a legitimate skill may reference URLs for

documentation purposes. Use the report as a starting point for human review, not as a definitive verdict.

• This skill does NOT prevent execution of other skills — it only reports findings.

Enforcement requires a human decision to remove or quarantine flagged skills.

Running the Scanner Manually

python3 ~/.claude/skills/security-monitor/scripts/scan_skills.py \
    --paths ~/.claude/skills .claude/skills \
    --report-format text

---

Part of the Claude Skills Defense Suite.

Execution

Run the security scanner now against all discoverable skill locations:

python3 "$(dirname "$0")/scripts/scan_skills.py" \
    --paths ~/.claude/skills ./.claude/skills \
    --report-format text

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: RationalEyes
• Source: RationalEyes/claude-skills-security-guide
• License: MIT

Install and usage instructions live in the source repository linked above.