JUMPING THE LINE: HOW MCP SERVERS CAN ATTACK YOU BEFORE YOU EVER USE THEM

Summary from the TLDR InfoSec newsletter

There is a critical vulnerability in the Model Context Protocol (MCP) known as "line jumping" where malicious servers can inject prompt instructions into tool descriptions before any tools are invoked. This bypasses security controls, allowing attackers to manipulate AI models to exfiltrate code, insert vulnerabilities, or suppress security alerts.