Cross-Layer Semantic Flow Reconstruction for Attack Detection in Agent Systems

The researchers propose AScope, an execution-aware framework that correlates application-layer agent semantics with kernel-layer audit events and reconstructs cross-layer semantic flows to detect attacks on agent systems. The framework connects fragmented operations into causal behavior trajectories and uses a supervised LLM to identify data-flow violations, control-flow deviations and intent inconsistencies. The paper evaluates on published AgentDojo trajectories and ten multi-agent scenarios with cross-layer telemetry, reporting node-level and path-level F1 of 85.3% and 66.7% respectively on the cross-layer dataset.