Skill Audit Methodology

Pipeline Overview

Our scanner runs automatically when a skill is submitted. Skills must pass all stages with a score of 70 or above and have zero critical or high findings to be approved.

Stage 1: Manifest Validation

Every skill must include a valid manifest with required metadata. The scanner checks:

• Required fields — name, slug, version, entrypoint, author
• Slug format — lowercase alphanumeric with hyphens (e.g. my-skill-name)
• Version format — strict semver (MAJOR.MINOR.PATCH)
• Entrypoint — must reference an existing file in the package

Stage 2: Network Permission Audit

Skills that request network access are subject to domain-level scrutiny:

• Domain limit — maximum 10 network domains per skill
• Wildcard detection — wildcard domains (e.g. *.example.com) are flagged as high severity
• Known exfiltration domains — checked against our threat intelligence feed
• Excessive scope — requesting more domains than the skill's functionality warrants

Stage 3: Source Code Scan

Static analysis powered by our supply-chain security engine scans for:

• Environment scanning — patterns that enumerate or exfiltrate environment variables
• Credential access — attempts to read SSH keys, tokens, or auth files
• Shell injection — unsafe use of exec, spawn, or template literals in commands
• Dependency risks — postinstall scripts, known malicious packages
• Package size — maximum 5MB per skill package

Stage 4: Sandbox Testing

Skills that pass stages 1–3 are executed in an isolated sandbox environment:

• Isolated container — runs in a seccomp-profiled container with no network access to production
• Permission enforcement — only declared permissions are granted
• Behavior monitoring — filesystem access, network calls, and process spawning are logged
• Resource limits — CPU, memory, and execution time are capped

Scoring System

Each skill starts at 100 points. Findings deduct points based on severity. Skills must score 70+ with no critical or high findings to be approved.

Approval Criteria

• Score of 70 or above after all deductions
• Zero critical findings
• Zero high findings
• All declared permissions must have a legitimate use case

Questions?

If you have questions about the audit process or need help fixing findings, contact us at marketplace@opensyber.cloud.