AIRS MLOps Lab Lab

Secure MLOps pipeline workshop: build, deploy, and secure ML models with AI Runtime Security, guided by Claude Code

Palo Alto Networks Internal

This lab is internal to Palo Alto Networks.

What You'll Build

The AIRS MLOps Lab is a hands-on workshop where you build, deploy, and secure a real ML pipeline. You fine-tune an open-source language model into a cybersecurity advisor, deploy it on Google Cloud, and then systematically secure the pipeline using Palo Alto Networks AI Runtime Security (AIRS).

This is not a passive tutorial. You work with Claude Code as your development partner and mentor. Claude has been configured specifically for this lab: it knows the codebase, paces its explanations, and uses Socratic questioning to build your understanding.

Prerequisites

Before starting the lab, ensure you have:

Torque — have access to creating your own projects prior to starting the lab
code.pan.run — access to code.pan.run
AIRS License — access to a Prisma AIRS tenant with Strata Cloud Manager credentials
Claude Code — installed and configured
GitHub Account — with gh CLI authenticated
Python 3.12+ — with uv for dependency management
HuggingFace Account — for exploring models and datasets

The Pipeline

You'll build a 3-gate CI/CD pipeline that scans ML models at every stage:

GATE 1: TRAIN GATE 2: PUBLISH GATE 3: DEPLOY ───────────── ─────────────── ────────────── Base Model Download Adapter Verify Manifest │ │ │ ▼ ▼ ▼ AIRS Scan Merge with Base AIRS Scan │ │ │ ▼ ▼ ▼ LoRA Training AIRS Scan Deploy Endpoint (Vertex AI) │ (vLLM + L4 GPU) │ ▼ │ ▼ Publish to GCS ▼ Create Manifest Deploy App (Cloud Run) └── auto_chain ──► └── auto_chain ──►

Gate 1 scans the base model before training, fine-tunes it with LoRA on Vertex AI, and creates a provenance manifest.

Gate 2 merges the LoRA adapter with the base model, scans the merged artifact, and publishes to GCS.

Gate 3 verifies the manifest chain, performs a final scan, deploys the model to a Vertex AI GPU endpoint with vLLM, and deploys the FastAPI app to Cloud Run.

The Architecture

The deployed application uses a decoupled architecture: the model runs on GPU infrastructure, and the app runs on lightweight infrastructure.

Browser │ POST /api/chat ▼ Cloud Run ── FastAPI (512Mi, 1 CPU) │ rawPredict (access token) ▼ Vertex AI ── rawPredict endpoint │ /v1/completions ▼ vLLM Server ── L4 GPU │ Qwen2.5-3B-Instruct (LoRA merged) ▼ completion ─► Vertex AI ─► Cloud Run ─► streamed reply to Browser

No model weights in the Cloud Run container. No GPU. No ML dependencies at runtime. Cloud Run handles the web layer; Vertex AI handles inference on dedicated GPU hardware.

The Three-Act Structure

The workshop is organized into three acts with a presentation break between Acts 1 and 2.

Act 1 — Build It (Modules 0-3)

Build a complete ML pipeline from scratch. By the end, you'll have a live cybersecurity advisor: trained, merged, published, and deployed. No security scans in place yet. That's intentional.

Module	Focus	Time
Module 0: Setup	Environment, GCP, GitHub, AIRS credentials	~30 min
Module 1: ML Fundamentals	HuggingFace, formats, datasets, platforms	~45 min
Module 2: Train Your Model	Gate 1, LoRA fine-tuning, Vertex AI	~30 min + wait
Module 3: Deploy & Serve	Gate 2+3, merge, publish, deploy, live app	~30 min + wait

Presentation Break

Instructor-led session: AIRS value proposition, real-world attacks, customer scenarios.

Act 2 — Understand Security (Module 4)

Deep dive into AIRS Model Security. Set up access, run scans, explore policies.

Module	Focus	Time
Module 4: AIRS Deep Dive	SCM, SDK, scanning, security groups, HF integration	~1-1.5 hr

Act 3 — Secure It (Modules 5-7)

Secure the pipeline you built. Explore what AIRS catches and what it doesn't.

Module	Focus	Time
Module 5: Integrate AIRS	Pipeline scanning, manifest verification, labeling	~1-1.5 hr
Module 6: The Threat Zoo	Pickle bombs, Keras traps, format risks	~1 hr
Module 7: Gaps & Poisoning	Data poisoning, behavioral backdoors, defense in depth	~45 min-1 hr

Project Structure

prisma-airs-mlops-lab/
├── .github/
│   ├── workflows/              # CI/CD pipeline (3 gates + app deploy)
│   └── pipeline-config.yaml    # Your GCP project and bucket config
├── src/airs_mlops_lab/
│   └── serving/                # FastAPI app + Vertex AI inference client
├── airs/
│   ├── scan_model.py           # AIRS scanning CLI
│   └── poisoning_demo/         # Data poisoning proof-of-concept
├── model-tuning/
│   ├── train_advisor.py        # LoRA fine-tuning script
│   └── merge_adapter.py        # Adapter merge for deployment
├── scripts/
│   └── manifest.py             # Model provenance tracking CLI
├── lab/
│   └── .progress.json          # Your progress (tracked automatically)
├── CLAUDE.md                   # Claude Code mentor configuration
└── Dockerfile                  # Cloud Run app (thin client, no model)

Next Step

Read How the Lab Works to understand the mentor, scoring, and pacing, then head to the Student Setup Guide to provision your GCP project via Torque, create your repo, and launch Claude Code.

Next: How the Lab Works →