guardscope: Measuring Which Actions Runtime Controls Check in LLM Agent Systems · Kentaroh Toyoda, Alex Leung (Vulcan Research, AIFT), Bhuvaneswari U (Standard Chartered) · arXiv, November 2026
A runtime control only sees the actions that pass through it
The authors list 15 paths an agent's action can take, from tool calls through a gateway to spawned background processes and direct network connections, and eight places a control can sit, from a gateway in front of the harness to the kernel. They then ran real controls, including the enforcement module of Microsoft's agent governance toolkit, agentgateway and the kernel-level ActPlane, with eBPF kernel tracing as ground truth. In their runs with a gateway in front of the harness plus the harness's own permission checks, both tool-call types were checked, but the eight direct-network, process-spawn, file-write and credential-access actions "take effect without any policy evaluation in every run." By their analysis, five of the 15 paths get past every placement short of the kernel, so stacking those controls adds detail about what is inspected but not coverage. Their workload exercised only 4 of the 15 paths; the rest rest on design analysis. They also measure coverage, not whether a policy decides correctly. The framework is released on GitHub.
Why it matters: If your agent controls live in a gateway or in the harness's permission prompts, actions the agent takes by shelling out or opening its own connections are not being checked. In the paper's runs, only a deployment that forced network traffic out through a controlled path stopped direct network calls, which is the job an organization-managed allowlist like OpenAI's is doing.