Execution isolation limits what a running process can access, change, or invoke. It can be implemented with boundaries such as containers, sandboxes, restricted operating-system permissions, or network controls. The goal is to contain a process even if its instructions or behavior are unsafe.
In cyber security, isolation reduces the damage a compromised or misbehaving process can cause. For AI agents, it can restrict access to files, data stores, commands, and external networks to only what a task requires. Isolation complements logging and output checks: those can reveal or assess behavior, while runtime restrictions can prevent unauthorized actions.