AgentSCOPE: Evaluating Contextual Privacy Across Agentic Workflows
arXiv 2603.04902·high signal
Introduces the Privacy Flow Graph grounded in Contextual Integrity theory to evaluate every boundary in an agentic pipeline, not just input/output. Tested 7 frontier LLMs across 62 multi-tool scenarios in 8 regulatory domains. Results: over 80% of scenarios contain intermediate privacy violations, and 24% of final outputs appear clean despite leaking data mid-pipeline. Most violations occur at the tool-response stage. Anyone building multi-tool agents needs to audit information flows, not just f