Security Researchers Warn Traditional SASE Cannot See What AI Agents Do Inside Encrypted Traffic

A newly published analysis argues that conventional Secure Access Service Edge (SASE) architectures have a critical blind spot: they can verify an encrypted connection to a large language model, but cannot inspect the intent of what an AI agent is doing inside that connection. The core concern is that an autonomous agent can invoke Model Context Protocol (MCP) tools to pull proprietary code or internal documents while a cloud proxy sees only valid, encrypted HTTPS traffic passing through.