LLM Security Scan GitHub Action

A GitHub Action that runs trusys-llm-scan in your CI to find LLM (Large Language Model) security issues in your code—prompt injection, unsafe model usage, data exposure, and framework-specific risks. Results can be reported as SARIF (for GitHub Security tab), JSON, or console output.

---

What this action does

1. Checks out your repo and sets up Python.
2. Installs trusys-llm-scan and Semgrep from PyPI.
3. Runs the scan on the paths you specify (default: whole repo).
4. Optionally uploads SARIF so findings show in the repository Security tab.
5. Optionally runs AI-based filtering or uploads results to your backend.

No config is required for a basic scan; add inputs only when you need custom paths, severity filters, AI, or upload.

Requirements: None in your repo. The action installs Python (via actions/setup-python) and the scanner (trusys-llm-scan) from PyPI. Your code only needs to be in a GitHub repository.

---

How to use the action (quick start)

1. Add a workflow file

In your repository, create or edit a file under .github/workflows/. For example:

• File: .github/workflows/llm-security-scan.yml

2. Minimum workflow (scan on push and pull requests)

Copy this into that file:

name: LLM Security Scan

on: [push, pull_request]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Run LLM Security Scan
        uses: spydra-tech/trusys-llm-security-scan-action@v1

That’s it. Every push and pull request will run the scanner. By default it:

• Scans the whole repository (.)
• Outputs in SARIF and uploads to GitHub so findings appear under Security → Code scanning
• Uses Python 3.11 (required by trusys-llm-scan; override with python-version if needed)

Where do I see results?
With the default format: 'sarif', the action uploads results to GitHub. Open your repo on GitHub → Security → Code scanning to see the findings. You can also read the job log for the full scan output.

3. Optional: pin to a specific version

• @v1 — use the latest v1.x (recommended for stability).
• @v1.0.0 — use an exact version.
• @main — use the latest from the default branch (may change without notice).

---

Features

• 🔍 Comprehensive Scanning: Detects LLM-specific vulnerabilities across multiple frameworks (OpenAI, Anthropic, Cohere, Langchain, LlamaIndex, Hugging Face, Azure, AWS Bedrock)
• 🤖 AI-Powered Analysis: Optional AI filtering to reduce false positives and enhance remediation guidance
• 📊 Multiple Output Formats: Supports console, JSON, and SARIF formats
• 🔗 Backend Integration: Optional upload of results to your backend API
• ⚡ Fast & Reliable: Built on Semgrep for efficient static analysis

---

Common configurations

What you want	Add this under with:
Scan only some folders	paths: 'src,lib'
Only high/critical findings	severity: 'critical,high'
Ignore tests and deps	exclude: '**/test/**,**/node_modules/**'
Different Python	python-version: '3.11'
JSON instead of SARIF	format: 'json'
AI false-positive filtering	enable-ai-filter: 'true' and ai-api-key: ${{ secrets.OPENAI_API_KEY }}
Send results to your backend	upload-endpoint: 'https://...', application-id: '...', api-key: ${{ secrets.API_KEY }}

Full list of inputs is in the Inputs table below.

---

Usage examples

Basic (defaults: whole repo, SARIF, Security tab)

name: LLM Security Scan

on: [push, pull_request]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Run LLM Security Scan
        uses: spydra-tech/trusys-llm-security-scan-action@v1

Paths, severity, and exclusions

- uses: spydra-tech/trusys-llm-security-scan-action@v1
  with:
    paths: 'src,lib'
    severity: 'high,critical'
    exclude: '**/test/**,**/node_modules/**'
    format: 'sarif'

With AI-powered false-positive filtering

Requires an API key in GitHub Secrets (e.g. OPENAI_API_KEY or ANTHROPIC_API_KEY).

- uses: spydra-tech/trusys-llm-security-scan-action@v1
  with:
    paths: 'src,lib'
    severity: 'high,critical'
    exclude: '**/test/**,**/node_modules/**'
    format: 'sarif'
    enable-ai-filter: 'true'
    ai-provider: 'openai'
    ai-model: 'gpt-4'
    ai-api-key: ${{ secrets.OPENAI_API_KEY }}
    ai-confidence-threshold: '0.8'
    ai-max-findings: '100'

With backend upload

name: LLM Security Scan

on: [push, pull_request]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      
      - name: Run LLM Security Scan and Upload
        uses: spydra-tech/trusys-llm-security-scan-action@v1
        with:
          format: 'sarif'
          upload-endpoint: 'https://api.example.com/scans'
          application-id: ${{ secrets.APPLICATION_ID }}
          api-key: ${{ secrets.API_KEY }}

Inputs

All inputs are optional. Omit any you don’t need.

Input	Description	Required	Default
paths	Comma-separated list of paths to scan	No	.
severity	Filter findings by severity (critical, high, medium, low, info)	No	``
exclude	Comma-separated list of patterns to exclude	No	``
include	Comma-separated list of patterns to include	No	``
python-version	Python version to use (3.11+ required by scanner)	No	3.11
format	Output format (console, json, sarif)	No	sarif
output	Output file path	No	``
enable-ai-filter	Enable AI-powered false positive filtering	No	false
ai-provider	AI provider (openai, anthropic)	No	openai
ai-model	AI model to use (e.g., gpt-4, claude-3-opus)	No	gpt-4
ai-api-key	API key for AI provider (use secrets)	No	``
ai-confidence-threshold	Minimum confidence threshold for AI analysis (0.0-1.0)	No	0.7
ai-max-findings	Maximum number of findings to analyze with AI	No	``
upload-endpoint	Backend API endpoint for uploading scan results	No	``
application-id	Application ID for backend upload	No	``
api-key	API key for backend upload (use secrets)	No	``

Outputs

Output	Description
scan-findings	Number of findings detected
sarif-path	Path to SARIF output file

Examples

Scan Specific Directories

- uses: spydra-tech/trusys-llm-security-scan-action@v1
  with:
    paths: 'src,lib,tests'

Filter by Severity

- uses: spydra-tech/trusys-llm-security-scan-action@v1
  with:
    severity: 'critical,high'

Exclude Patterns

- uses: spydra-tech/trusys-llm-security-scan-action@v1
  with:
    exclude: '**/test/**,**/node_modules/**,**/*.test.py'

AI Analysis with Custom Threshold

- uses: spydra-tech/trusys-llm-security-scan-action@v1
  with:
    enable-ai-filter: 'true'
    ai-provider: 'anthropic'
    ai-model: 'claude-3-opus'
    ai-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
    ai-confidence-threshold: '0.85'
    ai-max-findings: '50'

Upload to Backend

- uses: spydra-tech/trusys-llm-security-scan-action@v1
  with:
    format: 'json'
    upload-endpoint: 'https://api.example.com/api/scans'
    application-id: 'my-app-123'
    api-key: ${{ secrets.BACKEND_API_KEY }}

Security Best Practices

1. Never commit API keys: Always use GitHub Secrets for sensitive values like ai-api-key and api-key
2. Use least privilege: Only grant necessary permissions to the action
3. Review findings: Regularly review scan results and update your code accordingly
4. Keep updated: Use the latest version of the action for security updates

Supported Frameworks

• OpenAI
• Anthropic (Claude)
• Cohere
• Langchain
• LlamaIndex
• Hugging Face
• Azure OpenAI
• AWS Bedrock

Requirements

• In your repo: Nothing. The action installs Python (3.11 by default; trusys-llm-scan requires 3.11+) and the scanner (trusys-llm-scan) from PyPI.
• GitHub: A repository with Actions enabled; add the workflow file under .github/workflows/.

Troubleshooting

Scanner not found

Ensure Python is properly set up. The action uses actions/setup-python@v5 by default.

AI analysis not working

• Verify your API key is correctly set in GitHub Secrets
• Check that the AI provider and model are supported
• Review the action logs for detailed error messages

Upload failing

• Verify all required inputs (upload-endpoint, application-id, api-key) are provided
• Check that your backend API is accessible from GitHub Actions runners
• Review network connectivity and authentication

License

MIT License

Support

For issues and questions, please open an issue in the repository.