Security Pipeline

​

Overview

Every skill and agent submission passes through a multi-layer security pipeline before listing. No manual review is required for most submissions — the pipeline runs automatically.

​

Pipeline Stages

The pipeline runs 9 layers in sequence. Layers that don’t apply (e.g., dependency scanning on a pure-markdown skill) are skipped gracefully.

​

1. File Type Validation

Only allowed file types pass through. Disallowed extensions (executables, compiled binaries) are rejected immediately.

​

2. Pattern Scanning

Checks for dangerous patterns: shell commands, environment variable harvesting, obfuscation techniques, and known attack signatures.

​

3. ClamAV Signature Scan

A local ClamAV malware signature scan catches known malware before spending external API quota. Runs in ~200ms.

​

4. Dependency Vulnerability Scan (Snyk)

Scans dependency manifests (requirements.txt, package.json, pyproject.toml) for known CVEs using Snyk. Skipped if no manifests are present.

​

5. Supply Chain Analysis — Python (GuardDog)

GuardDog detects Python supply chain attacks that vulnerability scanners miss: typosquatting, dependency confusion, and install-time code execution.

​

6. Supply Chain Analysis — npm (Socket.dev)

Socket.dev analyzes npm dependencies for supply chain risks including install scripts, network access, and obfuscated code.

​

7. VirusTotal Malware Scan

The full package is uploaded to VirusTotal for scanning by 70+ antivirus engines. Results are polled with linear backoff.

​

8. AI Security Review

An AI-powered semantic analysis checks for:

• Prompt injection — Attempts to override Claude’s instructions
• Authentication flaws — Improper credential handling
• Data exposure — Leaking user data or API keys
• Privilege escalation — Requesting unnecessary permissions

​

9. Agent-Specific Checks

Submissions of type agent receive additional scrutiny:

• Tool access audit — Which tools does the agent request?
• Instruction analysis — Are instructions appropriate for the declared purpose?
• Trust level scoring — Overall risk assessment based on capabilities requested

​

Outcomes

​

Timing

Most submissions complete the full pipeline within minutes. Complex packages or high-traffic periods may take longer.

​

For Creators

​

Avoiding Rejections

• Do not include shell commands that modify system files
• Do not access environment variables beyond what your skill needs
• Do not attempt to override Claude’s safety instructions
• Keep file types to text, markdown, JSON, and standard code files

​

Appealing a Block

If your submission is blocked and you believe the finding is a false positive, update your package to address the findings and resubmit. Each resubmission goes through the full pipeline again.

​

External Skills (ClawHub and More)

External skills from registries like ClawHub go through the same full security pipeline. The difference is timing:

• Native skills — scanned at publish time, before they appear in the marketplace
• External skills — scanned at install time, with results cached so subsequent users get instant results

If an author updates a skill (changing its files), the cache is invalidated and a fresh scan runs automatically on the next install.

​

For Buyers

​

Reading Security Status

Every skill in the marketplace displays its security status:

• Pass — No issues detected
• Warn — Minor concerns flagged; review the badge details before installing

Skills with a Block status are never listed.

​

Reporting Concerns

If you discover a security issue with an installed skill, contact support at [email protected] with the skill slug and a description of the concern.

​

Need Help?