guardscope checks which AI agent actions security controls inspect

Researchers released guardscope, a framework for measuring which AI agent actions security controls actually inspect, according to an arXiv preprint. It compares control decision records with operating-system kernel records to classify executed actions as checked, bypassed, or prevented. The framework first examines control source code to predict coverage across 15 action paths, then tests a fixed workload in one of eight deployment designs. Researchers report experiments with two agent harnesses and controls including Microsoft's agent governance toolkit and agentgateway. The framework measures inspection coverage; an agent can take actions through paths beyond the tool calls a gateway checks.