A mechanism where the model outputs structured instructions to invoke external functions, APIs, or services. Tool calling is a key inflection point for risk because it enables AI systems to take real-world actions; permissions, logging, and approval workflows are essential controls.
See: Agentic AI; Least privilege