Skip to content
    Gaincontrol

    MCP poisoning: when the server you trusted isn't the one answering today

    Gaincontrol Team··3 min read

    "The server you trusted" is doing more work in that sentence than it looks like. The Tool Schema Can Lie to You showed that a schema read at discovery isn't a promise the server has to keep by the time it's actually called. This is the layer above that problem: the server itself isn't a promise either. An agent doesn't verify an operator, a company, or a person when it connects to an MCP server. It verifies a name, pointed at an address, that someone recorded in a config file or an allowlist at some point in the past. Whatever answers at that name today is "the server you trusted," whether or not it's the same thing that earned the trust.

    What gets reviewed is a moment, not an operator

    A new MCP integration usually gets looked at once: someone checks what the server does, decides it's fine, and adds its name to an allowlist. MCP Server Security already named the general version of this gap — a tool's name and description are just strings the server chooses to return, and nothing stops a compromised or malicious server from serving something else later. Server-level poisoning is the same gap one level up the stack: the entry in the allowlist doesn't track an operator's identity, a certificate, or a signed release. It tracks a string. If whatever controls that name or address changes — a registry listing gets tampered with, a lookalike entry slips in next to the real one, a legitimate server's infrastructure is compromised outright — the allowlist entry still matches. Nothing about the review that happened on day one checks whether it's still describing the thing calling itself by that name on day two hundred.

    Trusting the name is trusting whoever currently holds it

    This is the same shape as every supply-chain trust problem that predates agents: a package name, a domain, a vendor listing, all of them are only as trustworthy as whoever currently controls the thing behind the label. The Model Context Protocol's own tool model says clients must treat tool annotations as untrusted unless they come from a trusted server — but it leaves "trusted server" undefined, because the protocol has no mechanism for an agent to verify that the server answering right now is the one that earned that label. That's a gap the protocol was never going to close on its own; it describes how agents and servers talk, not who's allowed to answer to a given name.

    An allowlist that checks a name once is trusting a server that may no longer be the one it approved.

    Re-verify the server, not just the one-time review

    The fix isn't a better onboarding review — a more thorough version of the same one-time check still only describes the server as it existed on the day it was done. What closes the gap is treating server identity as something to check on every call it's actually part of, the same way a call's tool name, schema, and arguments get checked, rather than something a config entry settles permanently. An allowlist entry has to mean "this specific, continuously verified server," not "a name nobody's gotten around to re-checking."

    MCP made it easy to add a new integration. It didn't make "the server I added in March" the same claim as "the server answering right now," and treating those two as interchangeable is exactly the assumption server poisoning depends on.

    Sources

    • Tools. Model Context Protocol specification, version 2026-07-28. States that clients must treat tool annotations as untrusted unless they come from a trusted server, without defining what makes a server trusted or how that trust is verified on an ongoing basis.
    • MCP Security Notification: Tool Poisoning Attacks. Invariant Labs, April 2025. Describes compromised or malicious MCP servers serving different behavior than what was reviewed at onboarding, the same supply-chain trust gap this post names at the server level rather than the tool-schema level.

    Frequently asked questions

    What is MCP server poisoning?

    When the server actually answering to a name an agent already trusts isn't the one that earned that trust — through a tampered registry listing, a lookalike server, or a legitimate server's infrastructure being compromised outright.

    Why doesn't a one-time security review of an MCP server prevent this?

    Because the review only validates whatever was answering to that server's name on the day it happened. It has no way to confirm that the thing answering to that same name later is still the same operator, especially if the name or address itself was reassigned or hijacked after the review.

    How do you stop a trusted MCP server from being silently swapped?

    By checking server identity as part of every call it's involved in, not just once at onboarding — the same discipline already applied to the tool name, schema, and arguments on each call, rather than letting a config entry settle the question permanently.

    Zino authorizes every outgoing MCP call against a signed, continuously enforced policy before it leaves the agent — an unapproved or unrecognized server is denied by default, rather than trusted because a name in an allowlist matched.

    See how Zino checks MCP calls against policy, not a config entry →