Guard Action - AI Prompt Security

Detect prompt injection and PII in LLM prompts directly in your CI/CD pipeline.

Add one step to your GitHub Actions workflow to automatically scan AI prompts for security risks before they reach production.

Powered by @ny-squared/guard -- zero external API calls, fully offline, sub-millisecond performance.

---

Quick Start

- name: Scan AI prompt
  uses: ny-squared/guard-action@v1
  with:
    prompt: ${{ env.LLM_PROMPT }}

That's it. If a prompt injection or PII is detected above the default threshold (0.7), the workflow fails automatically.

---

What It Detects

Threat Type	Examples
Prompt Injection	"Ignore all previous instructions", "[SYSTEM] override", jailbreak patterns
PII Leakage	Email addresses, phone numbers, SSNs, credit card numbers

---

Inputs

Input	Required	Default	Description
prompt	Yes	-	The LLM prompt text to scan
threshold	No	0.7	Risk score threshold (0.0-1.0). Fails if exceeded
fail_on_block	No	true	Fail the workflow if prompt is blocked
api_key	No	-	PromptSentry API key (optional, for advanced cloud features)

Outputs

Output	Description
risk_score	Risk score from 0.0 to 1.0
blocked	Whether the prompt was blocked (true/false)
risk_type	Type of risk detected (injection/pii/none)
details	JSON string with full scan details

---

Usage Examples

Basic: Block dangerous prompts

name: AI Security Check
on: [push, pull_request]

jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - name: Scan prompt for injection
        uses: ny-squared/guard-action@v1
        with:
          prompt: "Translate this text to French: Hello world"
          threshold: '0.7'

Custom threshold with PII check

- name: Check for PII in prompt
  uses: ny-squared/guard-action@v1
  with:
    prompt: ${{ steps.build-prompt.outputs.text }}
    threshold: '0.5'
    fail_on_block: 'true'

Get score without failing (custom logic)

- name: Get risk score
  id: guard
  uses: ny-squared/guard-action@v1
  with:
    prompt: ${{ env.PROMPT }}
    fail_on_block: 'false'

- name: Custom risk handling
  if: steps.guard.outputs.blocked == 'true'
  run: |
    echo "Risk Score: ${{ steps.guard.outputs.risk_score }}"
    echo "Risk Type: ${{ steps.guard.outputs.risk_type }}"
    echo "Details: ${{ steps.guard.outputs.details }}"
    # Add custom alerting, logging, or notification logic here

PR review: Scan prompt templates on change

name: Prompt Security Review
on:
  pull_request:
    paths:
      - 'prompts/**'
      - 'src/prompts/**'

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

      - name: Read prompt template
        id: read-prompt
        run: |
          PROMPT=$(cat prompts/system-prompt.txt)
          echo "text=$PROMPT" >> $GITHUB_OUTPUT

      - name: Scan prompt template
        uses: ny-squared/guard-action@v1
        with:
          prompt: ${{ steps.read-prompt.outputs.text }}
          threshold: '0.6'

---

How It Works

1. Your workflow passes a prompt string to the action
2. @ny-squared/guard scans it locally (no external API calls) for injection patterns and PII
3. The action outputs the risk score, threat type, and block status
4. If the risk score exceeds your threshold and fail_on_block is true, the workflow fails

All scanning happens inside the GitHub Actions runner. No data leaves your CI environment.

---

Advanced: PromptSentry Cloud

For advanced features (cloud-based scanning, audit logs, team dashboard), provide your PromptSentry API key:

- name: Scan with PromptSentry Cloud
  uses: ny-squared/guard-action@v1
  with:
    prompt: ${{ env.PROMPT }}
    api_key: ${{ secrets.PROMPTSENTRY_API_KEY }}

Get your API key at app.trypromptguard.com.

---

Related

• @ny-squared/guard - The underlying SDK
• PromptSentry Dashboard - Web-based threat monitoring
• AI Security Benchmark - Test your LLM security

---

License

MIT