Skillspector Docker Image for Harness STO

Docker image that wraps NVIDIA Skillspector for use as a Harness STO (Security Testing Orchestration) step. Scans AI agent skills for security vulnerabilities and outputs results in SARIF format.

What is Skillspector?

Skillspector is a security scanner for AI agent skills (Claude Code, Cursor, etc.) that detects:

• Prompt injection vulnerabilities
• Data exfiltration risks
• Privilege escalation attempts
• Supply chain security issues
• 64+ vulnerability patterns across 16 categories

Quick Start

Build the Image

docker build -t skillspector-sto:latest .

Run Locally

Scan a directory containing skill files (SKILL.md):

docker run --rm \
  -v $(pwd)/my-skills:/workspace \
  -v $(pwd)/output:/output \
  skillspector-sto:latest

Configuration Parameters

Configure via environment variables:

Variable	Description	Default
SCAN_PATH	Path to scan for skill files	/workspace
OUTPUT_PATH	SARIF output file path	/output/results.sarif
VERBOSE	Enable verbose logging	false
NO_LLM	Disable LLM analysis (faster)	false
SKILLSPECTOR_LOG_LEVEL	Log level (DEBUG/INFO/WARNING/ERROR)	INFO

LLM Provider Configuration (Optional)

For enhanced semantic analysis, configure an LLM provider:

Variable	Description
SKILLSPECTOR_PROVIDER	Provider (openai/anthropic/nv_build)
OPENAI_API_KEY	OpenAI API key
ANTHROPIC_API_KEY	Anthropic API key
NVIDIA_INFERENCE_KEY	NVIDIA Inference API key
SKILLSPECTOR_MODEL	Override default model

Harness STO Integration

Pipeline Setup

Your Harness pipeline should:

1. Clone Skillspector to /skillspector/ (you already have this configured)
2. Add a Run step before Build to copy Skillspector into the build context:

- step:
    type: Run
    name: Copy Skillspector to Build Context
    identifier: copy_skillspector
    spec:
      shell: Sh
      command: |-
        # Copy Skillspector source into build context
        cp -r /skillspector/* .
        
        # Verify pyproject.toml exists
        test -f pyproject.toml && echo "✓ pyproject.toml found" || echo "✗ pyproject.toml missing"

Then your BuildAndPushDockerRegistry step will build from . which now contains Skillspector's source files.

Local Testing

Test the Docker image with a sample skill:

# Create a test skill
mkdir -p test-skills
cat > test-skills/SKILL.md << 'EOF'
---
name: test-skill
---

# Test Skill

This is a test skill for scanning.
EOF

# Run scan
docker run --rm \
  -v $(pwd)/test-skills:/workspace \
  -v $(pwd)/output:/output \
  -e VERBOSE=true \
  skillspector-sto:latest

# Check results
cat output/results.sarif

SARIF Output

The tool outputs SARIF 2.1.0 format, which Harness STO can ingest to:

• Display vulnerability findings in the UI
• Fail builds based on severity thresholds
• Track security posture over time
• Generate compliance reports

Troubleshooting

No findings in SARIF

• Ensure your repository contains SKILL.md files
• Check the SCAN_PATH is correctly mounted
• Enable VERBOSE=true to see scan details

Build failures

• Requires Python 3.12+ base image
• Ensure git is available during build
• Check network access to GitHub for cloning Skillspector

Performance

• Use NO_LLM=true for faster scans (static analysis only)
• LLM analysis provides deeper semantic vulnerability detection
• Consider resource limits in Harness (1Gi memory recommended)

License

This wrapper is provided as-is. Skillspector is licensed under Apache 2.0 by NVIDIA.