Skill Scanner
Use snyk/agent-scan to detect security risks in agent components.
Quick Scan
# Scan all skills on the machine
uvx snyk-agent-scan@latest --skills
# Scan MCP servers (default behavior)
uvx snyk-agent-scan@latest
# Scan with verbose output
uvx snyk-agent-scan@latest --skills --verbose
# Output JSON for automation
uvx snyk-agent-scan@latest --skills --json
What It Detects
For Skills
- Prompt Injection (E004) - Malicious instructions hidden in prompts
- Malware Payloads (E006) - Harmful code disguised as content
- Untrusted Content (W011) - Potentially unsafe external data
- Credential Handling (W007) - Improper secrets management
- Hardcoded Secrets (W008) - API keys or passwords in code
For MCP Servers
- Prompt Injection (E001)
- Tool Poisoning (E003)
- Tool Shadowing (E002)
- Toxic Flows (TF001)
- Rug Pull (W005) - Malicious skill replacement
Workflow
- Before installing a new skill → Run a scan first
- After scanning → Review any E001/E003/E004/E006 issues (high severity)
- Low severity warnings (W005-W008) → Decide based on your risk tolerance
Interpreting Results
| Prefix | Severity | Action |
|---|---|---|
| E | High | Fix or avoid the skill |
| W | Medium/Low | Review and decide |
| TF | High | Toxic flow detected |
Common Issues
If uvx is not found, install uv first:
# macOS
brew install uv
# Linux
curl -LsSf https://astral.sh/uv/install.sh | sh
OpenClaw Skills Location
OpenClaw skills are typically stored at:
- Global:
~/.openclaw/skills/ - Workspace:
<project>/skills/
To scan a custom path, pass it directly:
uvx snyk-agent-scan@latest ~/.openclaw/skills/
Output Example
The scan will show:
- File path of the issue
- Risk type and description
- Severity level (E/W/TF)
- Recommended fix
Review the full report at: https://github.com/snyk/agent-scan/blob/main/docs/issue-codes.md