A new preprint on arXiv introduces CreepScope, an approach to studying what the authors call privilege creep in tool-using agents. The core observation is that authorization checks at individual tool calls do not reveal how long earlier authority remains usable. In their framing, privilege creep occurs when a capability stays reachable after the workflow step that justified it has ended.

This shifts attention from static permission sets to the lifecycle of authority. Even if every call is authorized at the moment it is made, the window during which that authority can be exercised may extend far beyond the task that justified it. CreepScope is presented as a way to measure that post-use authorization exposure.

The preprint is the only source for this summary, so there are no independent results to compare. The work is identified as arXiv:2610.15988v1.