AI-201 Setup Guide

Prerequisite: AI-101 setup complete (Python 3.11+, OpenAI/Anthropic SDK, Burp Suite Community, Ollama, garak, PyRIT all installed and verified)

New installs for AI-201

Go 1.21+ (required for Lab 4: CVE-2025-9556)

# Linux / macOS
wget https://go.dev/dl/go1.21.13.linux-amd64.tar.gz
sudo tar -C /usr/local -xzf go1.21.13.linux-amd64.tar.gz
echo 'export PATH=$PATH:/usr/local/go/bin' >> ~/.bashrc
source ~/.bashrc

# Verify
go version

If Go 1.21 is unavailable, any version >= 1.21 is acceptable. The CVE-2025-9556 reproduction requires a version of LangChainGo compatible with the vulnerable release (see Lab 4 for the exact commit hash).

LangChainGo (vulnerable version for Lab 4)

mkdir ~/ai201-labs && cd ~/ai201-labs
mkdir lab-4 && cd lab-4
go mod init lab4
go get github.com/tmc/langchaingo@v0.1.12  # pinned to vulnerable version

faster-whisper (Module 7.5 local ASR)

pip install faster-whisper
# Test the install
python3 -c "from faster_whisper import WhisperModel; print('faster-whisper OK')"

Ollama: LLaVA and Whisper models (Module 7.5)

# LLaVA for visual prompt injection
ollama pull llava:7b

# Small Whisper-compatible model via Ollama
ollama pull whisper:base

# Verify
ollama list

Note: llava:7b is ~4.7 GB. Pull ahead of Module 7.5 to avoid waiting during lab time.

fickling + pickletools (Module 3)

pip install fickling
# pickletools is in Python stdlib; no separate install needed
python3 -c "import pickletools; import fickling; print('pickle tools OK')"

HarmBench evaluation framework (Module 4.5)

HarmBench runs on a cloud GPU. Setup is done in the Colab/Kaggle environment at lab time; no local install required. Verify access to Google Colab with GPU runtime before Module 4.5.

Promptfoo (Module 2 regression testing)

npm install -g promptfoo
# Verify
promptfoo --version

MITRE ATLAS Navigator (Module 1)

Browser-based tool; no install. URL: https://mitre-atlas.github.io/atlas-navigator/

Alternatively, the offline JSON data files are at github.com/mitre-atlas/atlas-data.

virtus-llm-owasp DVLA testbed (Module 2)

git clone https://github.com/virtus-cybersecurity/virtus-llm-owasp ~/virtus-dvla
cd ~/virtus-dvla
pip install -r requirements.txt

Verify you can start the DVLA locally: follow the README in the repo for the startup command. The academy's published L3-regression baseline is in baselines/l3-regression-v2.json.

Verify the full AI-201 environment

Run this before Module 1:

# Python ecosystem
python3 -c "import openai, anthropic, langchain, fickling; print('Python AI-201 stack OK')"

# Go
go version | grep -E "go1\.(2[1-9]|[3-9][0-9])"

# Ollama models
ollama list | grep -E "llava|whisper"

# DVLA
test -f ~/virtus-dvla/requirements.txt && echo "DVLA repo cloned OK"

# Promptfoo
promptfoo --version

All checks should pass. Bring blockers to the first lab session.

Tool Journal: AI-201 originating entries

Record the first entry for each tool below when you complete the corresponding module. See AI-101 §SETUP for the Toolchain Diary format.