Longer-form research on the problems I keep running into: agentic kill-chains, non-human identity, prompt injection as an output problem, and reading regulation into real controls.