AI alignment, mechanistic interpretability, model behavior, high-stakes agentic systems