AI Security Homelab

An isolated homelab for attacking and hardening AI/ML infrastructure: attack, harden, attack again, document.

Status
Active
Date
Context
Personal research
Scope
Own lab
  • Code assisted
  • Writing manual
What do these mean?

Overview

An isolated homelab for attacking and hardening AI/ML infrastructure: attack, harden, attack again, document. The stack runs locally under Docker Compose: Ollama and LiteLLM for inference, Qdrant for vectors, MinIO and MLflow for artifacts and tracking, FastAPI and LangGraph for the agent surface, Llama Guard for guardrails, and Wazuh for detection.

The seven-part series below is the lab’s notebook: each part takes one attack surface, breaks it, hardens it, and records what the detection layer actually saw.

Planned, not started yet: a model extraction / membership inference attack surface, and a supply-chain check for malicious pickle files.

Writing about this project