Researchers introduced guardscope in a paper submitted on 11 November 2026, proposing a way to measure which actions an LLM agent’s runtime controls can check. The authors say a control sees only actions that cross its position in a system, while an agent may act through paths beyond the tool-call channel.[1]
The paper describes two stages. First, static analysis predicts which of 15 agent action paths a control can check from its source. Second, an empirical stage runs a fixed workload on a deployment using one of eight topology designs. It reconciles the control’s decision records with kernel-level ground truth and classifies executed actions as checked, bypassed or prevented.[1]
The authors say a control developer could use the framework to publish which paths its product inspects under a kernel-verified protocol. They report experiments with two agent harnesses and real runtime controls, including Microsoft’s agent governance toolkit and agentgateway. The saved abstract does not report the experiments’ results or identify a recommended fixed version.[1]
[1]: https://arxiv.org/abs/2611.15223
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
125cbab9377b1f043168fca891c236a80275b314ef571901293199d9731a5bf6