Notes on AI security, platform engineering, and zero-trust built from Markdown at deploy time.
2026-07-28
MLX LoRA on Apple Silicon: build a dataset, score a baseline, train an adapter, and check whether it actually learned anything — without mistaking loss for correctness.
./read-post
2026-06-02
What happens when you hide an attack inside an AI agent's tool output and why the system prompt won't save you. A hands-on look at prompt injection and OPA-enforced tool authorization for local LLM agents. Part 1 of the AgentFence series.
./read-post