clawshield
# ClawShield
Scan a skill directory without executing it and return a risk level that can be enforced in review or CI.
## When to use
- You want a fast static review before installing or publishing a skill.
- You need machine-readable findings for CI or release gates.
- You want a narrow ruleset aimed at common high-risk supply-chain patterns.
## Command
```bash
node {baseDir}/bin/clawshield.js scan /path/to/skill --format table
node {baseDir}/bin/clawshield.js scan /path/to/skill --format json
node {baseDir}/bin/clawshield.js scan /path/to/skill --format sarif > clawshield.sarif
node {baseDir}/bin/clawshield.js scan /path/to/skill --format table --fail-on caution
```
## Rules
| Rule ID | Severity | Description |
| --- | --- | --- |
| CS001_CURL_PIPE_SH | high | `curl` or `wget` piped directly into a shell |
| CS002_OBFUSCATED_EXEC | high | obfuscated or dynamic execution such as `eval`, `new Function`, or base64 decode flows |
| CS003_SUSPICIOUS_CALLBACK | medium | suspicious outbound callback endpoints such as raw IPs, ngrok, or webhook collectors |
| CS004_SOCIAL_ENGINEERING_PROMPT | medium | instructions that pressure users to bypass safety controls |
| CS005_SHELL_WRAPPER_EXEC | high | `bash -c` wrappers that hide remote execution |
## Risk levels
- **Safe**: no findings after suppressions
- **Caution**: one or more medium-severity findings
- **Avoid**: one or more high-severity findings
## Suppressions
Create `.clawshield-suppressions.json` in the target skill directory:
```json
[
{
"ruleId": "CS001_CURL_PIPE_SH",
"file": "install.sh",
"line": 15,
"justification": "Reviewed manually; uses a pinned artifact with signature verification."
}
]
```
Suppressions without justification are ignored.
## CI example
```yaml
- run: node {baseDir}/bin/clawshield.js scan . --format sarif --fail-on caution
```
## Boundaries
- ClawShield is a static scanner. It does not sandbox or execute the target skill.
- The rule set is intentionally narrow and should be treated as a high-signal first pass, not a full security audit.
标签
skill
ai
