GAAP: Deterministic Privacy Enforcement for AI Agents Using Information Flow Control — No Trust in Agent Required

arXiv·medium signal

Published April 21 on arXiv by Stanley, Verma, Tsai, Kallas, and Kumar, GAAP (Guaranteed Accounting for Agent Privacy) introduces an AI agent execution environment that enforces user-defined data sharing permissions deterministically without trusting the agent, the model, or the prompt to be attack-free. The system uses information flow control techniques to monitor how agents access and use private data across tasks, blocking unauthorized disclosure attempts without significantly reducing agent functionality. This addresses the core trust problem in agent deployment: how to give agents access to sensitive data while guaranteeing they can't leak it.