Trust & Safety

Every skill in the Agent Skill Exchange is backed by a real tool, repo, or package. New skills go through a discovery pipeline that requires real provenance before publishing.

How It Works

📋 Published               In the catalog — real tool, real provenance
     ↓
🛡️ Security Reviewed      Scanned for safety — highest trust

📋 Published

The skill is published in the catalog. To be published, a skill must:

• Be backed by a real tool, repo, or package
• Have a valid SKILL.md with required frontmatter
• Be assigned to a valid category and framework
• Have substantive content (100+ words with technical references)
• Pass deduplication checks against existing skills

🛡️ Security Reviewed

The skill's content has been scanned for potentially malicious patterns:

• No prompt injection attempts
• No data exfiltration patterns (unauthorized network calls)
• No destructive commands (rm -rf /, disk formatting, etc.)
• No credential harvesting instructions
• No obfuscated or encoded payloads
• No reverse shells or crypto mining

Checking a Skill's Status

Each skill's SKILL.md frontmatter includes a verification field:

---
title: "Example Skill"
verification: security_reviewed
---

Valid values:

Value	User-facing label	Meaning
listed	Published	In the catalog — backed by a real tool
security_reviewed	Security Reviewed	Content scanned for malicious patterns — safe to use

Executable Security Contract

Security review patterns are now represented as machine-readable data in patterns.json, with regression fixtures in fixtures/security/. CI runs:

python3 scripts/test_security_patterns.py
python3 scripts/security_scan.py skills --github-annotations --quiet

This keeps the trust layer testable instead of relying on documentation-only checks.

More Information

• Verification Criteria — detailed requirements for each tier
• Security Patterns — human-readable patterns checked during security review
• Browse Skills — explore the full catalog