| name | prisma-airs |
| description | Scan prompts, AI responses, and code for security threats using Prisma AIRS.
Auto-invoke when: checking content for prompt injection, detecting sensitive data (PII, credentials, secrets),
identifying malicious URLs, filtering toxic or harmful content, or validating AI-generated responses.
Auto-invoke when: user asks to "scan this", "check for sensitive data", "is this safe",
"check for injection", "review for security", or mentions DLP, PII, or credentials.
Auto-invoke when: generating code that handles user input or authentication.
|
| allowed-tools | ["Bash","Read","Write"] |
Prisma AIRS Security Scanner
Detect security threats in prompts, AI responses, and code using Palo Alto Networks Prisma AIRS.
What It Detects
- Prompt Injection - Attempts to manipulate AI behavior
- Data Loss Prevention (DLP) - PII, credentials, API keys, secrets
- Malicious URLs - Phishing, malware, command & control
- Toxic Content - Harmful, offensive, or inappropriate content
- Malicious Code - Exploits, malware patterns
Prerequisites
Environment variables required:
PRISMA_AIRS_API_KEY - API key from Strata Cloud Manager
PRISMA_AIRS_PROFILE_NAME - Security profile name
PRISMA_AIRS_URL - (Optional) Regional API endpoint
How to Pass Content
Choose the method based on content complexity:
Method 1: Heredoc (recommended for most content)
Use heredoc to avoid shell escaping issues:
python scripts/scan.py --type prompt <<'EOF'
Content to scan goes here.
Can include "quotes", newlines, and special chars.
EOF
Method 2: File (recommended for code or large content)
Write content to a temp file, then scan:
python scripts/scan.py --type code --file /tmp/scan-content.py
Method 3: Direct argument (simple content only)
Only use for short, simple strings without special characters:
python scripts/scan.py --type prompt --content "simple text here"
Scan Types
| Type | Use Case |
|---|
prompt | User input - check for injection, DLP, malicious content |
response | AI output - check for sensitive data leakage, toxic content |
code | Generated code - security vulnerabilities, malicious patterns |
conversation | Full context - prompt and response together |
Interpreting Results
| Action | Meaning |
|---|
allow | Safe to proceed |
alert | Review findings before proceeding |
block | Threat detected - do not proceed without remediation |
Workflow
- Extract content to scan from the conversation
- Choose appropriate scan type and method
- Run the scanner
- Check the
action field in the result
- If
block or alert: address the issue before proceeding
For threat category details, see references/threat-categories.md.