Sandboxing

The practice of running an AI system or agent in an isolated environment that limits its ability to interact with external systems, access sensitive resources, or take irreversible actions. Sandboxing is a critical containment control for agentic AI systems — by restricting what the agent can reach and do, it bounds the potential damage from errors, misuse, or adversarial exploitation. The more capable and autonomous the agent, the more important a well-enforced sandbox becomes.
See also: security boundary, agentic AI.