Notes on AI security, prompt injection, red-team experiments, and system tinkering.