Prisma AIRS SDK Example

Example scripts for batch-scanning prompts against Palo Alto Networks Prisma AIRS using the pan-aisecurity Python SDK.

Two scanning modes:

• main.py — synchronous, one prompt at a time with 0.5s delay
• main_threaded.py — multi-threaded with configurable worker pool (default 20)

Both scripts read from prompts.csv and write results back to the same file.

Prerequisites

• Python 3.11+
• Prisma AIRS API key and AI security profile name

Setup

git clone git@github.com:cdot65/prisma-airs-sdk-example.git
cd prisma-airs-sdk-example

Configure credentials

cp .env.example .env

Edit .env:

AIRS_API_KEY="your-api-key-here"
AIRS_AI_PROFILE="your-ai-profile-name-here"

Install dependencies

uv sync

Or with pip:

python -m venv .venv
source .venv/bin/activate
pip install pan-aisecurity python-dotenv

Prepare your prompts

cp prompts.example.csv prompts.csv

Edit prompts.csv to add your test prompts. Only three columns are required as input:

Column	Description
index	Row number (0-based)
prompt_preview	The prompt text to scan
expected	TRUE if the prompt should be blocked, FALSE if allowed

All other columns are populated by the scanner after running.

Usage

Synchronous scan

uv run main.py

Threaded scan

uv run main_threaded.py

Set AIRS_THREAD_COUNT to control worker pool size (minimum 20):

AIRS_THREAD_COUNT=50 uv run main_threaded.py

Output columns

After scanning, prompts.csv is updated with these result columns:

Column	Description
action	AIRS verdict: allow or block
category	AIRS classification: benign or malicious
scan_id	Scan identifier
report_id	Report identifier
agent	Agent security detection triggered
injection	Prompt injection detection triggered
dlp	Data loss prevention detection triggered
toxic_content	Toxic content detection triggered
malicious_code	Malicious code detection triggered
url_cats	URL filtering detection triggered
topic_violation	Topic violation detection triggered
match	PASS if result matches expected, FAIL otherwise, ERROR on failure
error_message	Error details if scan failed

Additional verdict detail columns (dlp_verdict, tc_verdict, overall_actions, etc.) are populated when report data is available.

Tests

uv run pytest

Linting

uv run ruff check --fix .
uv run ruff format .