What is an AI Safety?

In practice this spans several layers: alignment (does the model pursue the goals it was actually given), robustness (does it behave sanely on inputs it wasn't trained for), and control (can a human or system intervene, pause, or override it when something goes wrong). For agentic systems the stakes rise because the model isn't just generating text — it's executing actions with real-world side effects. Operational AI safety usually means layered defenses rather than a single switch: input validation, output filtering, scoped tool permissions, human-in-the-loop approval for high-risk actions, and continuous monitoring for drift or anomalous behavior. No single layer is trusted alone — the assumption is that any one control can fail, so the system stays safe only if multiple independent controls would all have to fail simultaneously. Safety work is never "done" — it's a continuous practice of red-teaming, incident review, and tightening controls as new failure modes surface, especially as models gain more autonomy and broader tool access over time.