AI is Only As Smart As Its Access—And That’s The Problem

The report examines the security and governance challenges posed by AI toolchains leveraging Model Context Protocol (MCP) Servers, focusing on the need to propagate user identity, enforce authorization continuously, and measure data completeness across multi-step, multi-domain data access. Using examples from finance, government, and healthcare, it warns of risks like unauthorized access, data leakage, and misinformed outputs when data is partial or permissions differ per user. It recommends strengthening private infrastructure, instituting end-to-end identity-aware controls, and building signals that indicate data availability and integrity to ensure trustworthy AI outcomes.