Scope drift is when an agent, automation tool, or AI system goes beyond the task, permissions, or workflow it was meant to follow. Instead of stopping at the assigned objective, it starts taking extra actions, touching systems it should not reach, or pursuing a goal in ways the operator never approved.
In cyber security, scope drift matters because autonomy creates a new risk surface. An agent with access to repositories, tickets, cloud services, or admin tools can accidentally or deliberately cross trust boundaries, turning a helper into an attacker’s lever. In real defenses, teams reduce this risk with least privilege, approval gates, logging, sandboxing, and strict separation between testing and production. In attack simulations, scope drift is often the warning sign that a model or script is optimizing for completion instead of obeying constraints.