Securing AI Agents, MCP & RAG in Production

A 2-Day Attack-Then-Defend Intensive — Tech Skills Transformations

Revision 1.0 — December 2026

A hands-on workshop where you break and then harden a realistic AI support agent — its RAG pipeline, its tools, its MCP gateway, and a coding agent — across the six layers of a practical security blueprint. Every lab is an attack followed by a defense, run in a GitHub Codespace. No security background required; Python literacy is assumed.

Getting started

1. Click the green Code button → Codespaces → Create codespace on main.
2. Wait 3–5 minutes for the environment to build. It creates a Python virtualenv, installs dependencies, and pulls the local model llama3.2:3b.
3. When setup finishes, open labs.md (it opens in preview by default) and start at Lab 1.

The lab model / API keys

Labs use a small local model (Ollama llama3.2:3b) by default — no account or key needed. Replies take 30s–2 min on a 4-core Codespace; that's expected.

To use a faster cloud model instead, set one of these in the terminal before running a lab (the labs auto-detect it):

export ANTHROPIC_API_KEY=...      # uses Claude
# or
export OPENAI_API_KEY=...         # uses GPT

System requirements

• A GitHub account (Codespaces free tier is sufficient for this course).
• A modern browser (Chrome recommended).
• Bonus lab only: a paid Claude account with Claude Code installed.

Merge steps and hover notes

Labs use a diff-and-merge workflow: code -d ../extra/<file>_complete.txt <skeleton>.py opens the completed reference on the left and your skeleton on the right. Merge the highlighted blocks left-to-right, then save.

merge-info.json at the repo root carries a short explanation for every block you will merge (28 of them across 15 files). With the Merge Info VS Code extension installed, hovering a block in the left pane shows that note. The labs work without the extension — the notes are optional depth.

Repository layout

foundations/  Lab 1  threat modeling
injection/    Lab 2  direct prompt injection
rag/          Lab 3  RAG poisoning + provenance filtering
data/         Lab 4  PII tokenization
guardrails/   Lab 5  guard pipeline + red-team harness
agents/       Lab 6  tool authorization policy
mcp/          Labs 7-8  MCP tool poisoning, scoped gateway
codingagent/  Lab 9  coding-agent injection
containment/  Lab 10 blast-radius containment
ops/          Lab 11 observability + CI security gate
capstone/     Labs 12-13 AI Act Article 50 controls, end-to-end hardening + scorecard
bonus-claudecode/  optional Claude Code hardening lab
extra/        completed reference code for diff-merge steps
merge-info.json  hover notes for every merge block
common/       shared model client used by all labs

Troubleshooting

• A lab fails with NotImplementedError: merge gap N — you ran a skeleton before completing its code -d merge step. Re-open the diff and merge the gap.
• Model replies are very slow — that's the local model. Set a cloud API key (above) for near-instant replies.
• ollama: connection refused — run bash scripts/startup_ollama.sh to restart the local model server.
• address already in use — a previous lab's process is still running; the labs use in-memory transport, so just start a fresh terminal.

License & attribution

For educational use only by the attendees of our workshops. © 2026 Tech Skills Transformations LLC and Brent C. Laster. All rights reserved. techskillstransformations.com