MCP tool poisoning means hiding malicious instructions in the description of a tool offered by a Model Context Protocol server or, more broadly, in what that tool returns to the model. The user only sees a name and a summary, but the model reads the full text and may comply: reading sensitive files, sending data out or hijacking another tool. Invariant Labs described the attack on 1 April 2025: a simple addition tool led the Cursor code editor to send the user's MCP configuration and private SSH key to the malicious server, and a server could change a tool's description after it had been approved. In 2026, the NSA published dedicated MCP security guidance on 20 May, and Microsoft reported on 30 June that it was observing these techniques against a growing range of enterprise agents.
In practice at Gensai
Gensai favours MCP servers whose publisher is identified, only enables the tools the use case needs and submits sensitive actions to human validation, with a log of tool calls.
What precautions should be taken before connecting an MCP server?
Treat it as third-party software: check who publishes it, read the full description of its tools, pin the installed version and review it again at every update. Only enable the tools you need, and have a person confirm sensitive actions. The MCP specification points the same way: clients must treat tool annotations as untrusted unless they come from a trusted server, and a person should always be able to deny a tool call.
Common mistake
Approving an MCP server once and for all. A tool can change after installation: the fake “postmark-mcp” email server, discovered on npm in September 2025, had worked normally for 15 versions before an update started quietly sending a blind copy of every email to an outside server. Every update of a third-party server deserves the same review as a change to the agent's instructions.
Same theme
Definition written with AI assistance, checked by Sébastien Lecocq, co-founder of Gensai.
Going beyond the definition?
From concept to project: Gensai builds custom, sovereign, GDPR-compliant AI solutions.
Related services: Chatbots and answer engines, Custom AI solutions