EchoLeak is a newly discovered zero-click AI vulnerability in Microsoft 365 Copilot that enables attackers to silently exfiltrate sensitive internal data without user interaction by exploiting a Large Language Model (LLM) Scope Violation. Identified by Aim Labs/Aim Security and patched server-side by Microsoft in May 2025, the flaw, tracked as CVE-2025-32711 (CVSS 9.3/10), leverages prompt injections embedded in seemingly benign emails that bypass security filters and are later retrieved into the AI’s context by the Retrieval-Augmented Generation (RAG) engine. This causes the LLM to extract privileged information and leak it via markdown-formatted image URLs or trusted Microsoft Teams and SharePoint links. Though not exploited in the wild, EchoLeak exemplifies a new class of vulnerabilities that weaponize helpful AI features into silent attack vectors. The attack’s automation potential and ability to operate without user behavior make it particularly dangerous in enterprise environments. The broader article also explores related threats to the Model Context Protocol (MCP), including Full-Schema Poisoning (FSP), advanced tool poisoning attacks (ATPA), and DNS rebinding techniques targeting localhost MCP servers via deprecated Server-Sent Events (SSE). These findings expose critical architectural blind spots in LLM-integrated systems, emphasizing the urgent need for improved prompt filtering, input scoping, output sanitization, and strict trust boundary enforcement in AI-agent design and infrastructure.