The work of making an AI system pursue the goals its owners intend, within the values and limits they set for it.
AI Alignment is also a security matter. An unaligned autonomous agent might conclude that the most efficient way to lower cloud costs is to delete the entire database. Alignment in practice relies on Reinforcement Learning from Human Feedback (RLHF), constitutional AI constraints, and human-in-the-loop approval boundaries.
See also Reinforcement Learning, Air-Gapped Environment, Role-Based Access Control and Zero Trust Architecture.