---
title: Skill Security Auditor
category: product
entity_type: skill
price: $15
canonical: https://forgehouse.ai/skills/skill-security-auditor/
lang: en
hreflang_alt: https://forgehouse.ai/tr/skiller/skill-security-auditor/
last_updated: 2026-06-20
---

# Skill Security Auditor

> Audit newly added or updated skill files for security threats.

An automated security gate that audits every new or updated skill before it enters your ecosystem. It runs a 7-check matrix: four red-level checks (hardcoded secrets, eval and dynamic code execution, external data exfiltration, and cross-skill modification) that auto-reject, plus three yellow-level checks that flag routing clashes, over-broad scope, and structural gaps for human review. One malicious skill can compromise the whole ecosystem, so it scans in seconds and fails secure.

## Use cases
- Auditing a skill before installing it
- Vetting third-party skills before import
- Self-auditing skills generated by other tools
- Catching hardcoded secrets and tokens in skill files
- Detecting skills that try to modify other skills
- Periodic re-audit of an entire skill library

## Benefits
- Block supply-chain attacks before a bad skill ever runs
- Fail secure: anything uncertain is rejected or escalated, never silently accepted
- Get a clear PASS / FAIL / REVIEW verdict with the specific reason
- Add new skills with confidence instead of manual line-by-line inspection

## What’s included
- Four auto-reject checks: secrets, eval, external exfiltration, cross-skill modification
- Three review checks: routing overlap, scope breadth, structural gaps
- Regex plus entropy detection for high-confidence secret scanning
- False-positive filters for educational examples and documentation links
- Edge-case handling for large files, binaries, symlinks, and multi-file skills
- Structured audit report with verdict, reasons, and audit trail logging

## Who it’s for
Platform owners and teams running a skill marketplace or internal library who need every addition vetted against supply-chain threats.

## How it runs
One malicious skill poisons the whole ecosystem, so every new file passes a gate first. Four RED checks reject hardcoded secrets, eval calls, outbound exfiltration and cross-skill tampering; three YELLOW checks route ambiguity to human review.
1. Read the full skill file, frontmatter, body and any bundled scripts; multi-file skills are audited file by file, symlinks resolved to their targets first.
2. Run the 4 RED checks in strict order: R1 hardcoded secrets via regex plus Shannon entropy analysis, R2 eval() and new Function() usage, R3 outbound POST/PUT/PATCH calls to external URLs, R4 attempts to modify other skills' directories. The first RED hit means automatic reject, the audit stops there.
3. Apply false positive filters before flagging anything: educational anti-pattern examples inside code blocks, regex definitions in detection tables and plain documentation links are recognized and skipped.
4. Run all 3 YELLOW checks regardless of outcome: Y1 routing keyword clash above 60 percent overlap with existing skills, Y2 over-broad description spanning 5 or more domains, Y3 missing structural sections. Each YELLOW finding goes to human review, never silent auto-reject.
5. Issue the verdict report, PASS, FAIL with the specific failing check, or REVIEW with a concrete recommendation, and write the audit trail with date and result so the decision is reconstructable later.

## FAQ
### We install third-party skills weekly, does this fit into that flow?
Yes, it is designed as a gate: it runs automatically through the /install-skill command and scans in seconds, so vetting third-party or generated skills before import is the primary workflow, not an occasional audit.

### What stops it from rubber-stamping a cleverly disguised malicious skill?
The matrix fails secure: four red-level checks (hardcoded secrets, eval, external exfiltration, cross-skill modification) auto-reject, secret scanning combines regex with entropy detection, and anything uncertain escalates to human review instead of passing silently.

### Does a PASS verdict mean the skill is completely safe?
No. It is a static 7-check gate, not a sandbox or runtime monitor. Logic bugs, prompt-level manipulation, or behavior that only emerges at execution time are out of scope; PASS means no known static threat patterns were found.

## Price
$15, one-time, no subscription. VAT included.

Related guide: [AI for application security](https://forgehouse.ai/guides/ai-application-security/)