Definition
A security flaw in a Model Context Protocol (MCP) server — the software that AI agents use to call external tools and services — where an attacker can gain access without providing valid credentials, or can forge authentication tokens that the server accepts as legitimate. Because MCP servers act as privileged gateways to email, file systems, APIs, and other sensitive resources, bypassing their authentication gives an attacker full control over every action the AI agent can perform. Multiple AI platforms have disclosed this class of vulnerability in 2026.
Why it matters
An MCP server with broken authentication is effectively an unlocked door to every system the AI agent is authorised to use — an unauthenticated attacker can impersonate the agent, read data, execute commands, and cover their tracks without ever touching a user's password.