When English becomes exploit code: The hidden risk inside MCP servers

A harmless-looking sentence can influence an agent's choices once it enters a model's context.