An MCP tool poisoning attack places malicious instructions in tool metadata that a model may use when deciding how to act. A description can be malicious when first connected, or it can change after approval. In the latter case, a client that does not review metadata changes may expose the agent to instructions the user never approved.
Post-approval substitution is one variant, often called a rug pull. The wider issue is whether untrusted tool text can redirect the agent’s use of otherwise legitimate permissions.
Inside an MCP tool poisoning attack: a finance workflow walkthrough
Microsoft Incident Response’s June 30, 2026 article describes a finance-workflow attack pattern, not a disclosed breach at a named company. An assistant uses a vendor database, email connector and third-party enrichment tool. After approval, the enrichment tool’s description changes while its visible name and summary remain the same.
In the scenario, the new description tells the agent to collect thirty unpaid invoices and attach a summary to an enrichment request. A routine supplier question then causes that unauthorized disclosure, even though the individual tool operations fit the agent’s existing access. The lesson is to review metadata changes and outbound payloads, not just whether a tool appears on an approved list.
Same tool, same name, new instructions
An installation review can miss later metadata changes. Microsoft describes the middle step as silent re-trust: where description changes do not trigger a re-approval workflow, “the updated instructions become active without additional review.”

In this scenario, the visible name and summary stay the same while the description changes.
Here is the part that makes this hard to reason about. A description is documentation that can also be supplied to the model to guide tool selection. In an arrangement where the client registers the whole tool list up front, that text is already in the model’s working context before any call is made, which is also the setup MCPTox tested below. A client that loads descriptions lazily or filters them changes that exposure, so it is worth knowing which kind yours is. Where the list is registered up front, a poisoned description does not need its own tool to be invoked to have an effect.
Text from different sources can coexist in model context without having the same intended authority. A client can preserve provenance, restrict tools and validate requests. The attack succeeds when those boundaries fail to prevent an untrusted instruction from changing the action.
You can check part of this on your own machine in about ten seconds. Open the tool permissions for any connector you have enabled. The panel I get lists readable names, a count and a permission control, and shows no description for any of them. The MCP specification allows an optional description field in a tools/list response. Querying it can reveal descriptions the panel does not display, where the server supplies them. The screenshot alone does not prove that every listed tool has one.

This panel lists fourteen tools but does not display full descriptions. Check the actual tool metadata and the scope of its permission controls separately.
Tool poisoning vs prompt injection: where the definitions split
Search for this term and you will find people describing three different attacks under one label. They are related, but if you are trying to decide what to check on Monday morning, the difference matters.
Poisoning the description before you install
This is the original version, demonstrated by Invariant Labs in April 2025: a tool that looks harmless, carrying hidden instructions from the moment you connect it. Our guide to MCP server security risks walks through their example in detail, so this post will not repeat it.
Poisoning the description after you approve it
Microsoft’s pattern above. The install was clean. The change came later, through an update channel that had no reviewer attached to it.
Poisoning the response at runtime
An OWASP community page also describes malicious instructions carried in tool responses. A client that passes such replies into model context without suitable checks can expose the agent to indirect prompt injection.
OWASP’s MCP Top 10 uses a broader tool-poisoning category covering tools, plugins and outputs. Its project roadmap still identifies the list as beta as checked in October 2026. The metadata and response routes share a trust-boundary problem, but they require checks at different points.
All three are prompt injection in the sense that matters: untrusted text reaching a model that treats it as instruction. What changes is the delivery route. If you have read direct vs indirect prompt injection, think of tool poisoning as indirect injection aimed at a channel that was never presented to you as content.
What MCPTox measured about tool poisoning
The intuitive assumption is that a smarter model would notice something was off. MCPTox reported a tendency in that direction for its tested agents, not a general law about intelligence.
The August 2025 version of MCPTox built 1,312 malicious test cases across 10 risk categories on top of 45 live MCP servers and 353 real tools, then ran them against 20 LLM agents. The highest attack success rate belonged to o1-mini, at 72.8 percent. That figure is one model’s result, the worst of the twenty, not an average across the ecosystem.
Two findings from that paper are more useful than the headline number. First, the researchers found that “more capable models are often more susceptible, as the attack exploits their superior instruction-following abilities.” Read the “often” as written: it is a tendency across twenty agent settings with a lot of variation inside them, not a law that says capability always makes things worse. Being good at following instructions is what the attack exploits. Second, refusal was close to absent: the best refusal rate in the group, from Claude-3.7-Sonnet, was still under 3 percent.
That is where the paper stops. My own reading, which is a conclusion rather than a finding of the research: waiting for models to get good enough to spot this is not a plan you can hold today. The benchmark is a single point in time, tested single-turn, and better clients and structured output validation may well move the numbers. What it does not support is doing nothing while that happens.
MCP security requirements bind servers and clients differently
It would be easy to read all of this as a protocol flaw. The spec is more honest than that, and reading the actual text is clarifying.
MCP does have a mechanism for tools changing. Servers that declare the listChanged capability SHOULD send a notification when their tool list changes. The spec also tells clients they MUST treat tool annotations as untrusted unless they come from a trusted server, says there SHOULD always be a human able to deny a tool invocation, and lists validating tool results before passing them to the model as something clients SHOULD do.
Notice which word is doing the work, and who it binds. The spec does place MUST obligations on servers: validate all tool inputs, implement proper access controls, rate limit invocations, sanitize outputs, and return results conforming to an output schema where one is declared. Those requirements matter for compliant implementations, but a malicious server may ignore them. Independent client controls remain necessary. On the client side, where your protection would have to come from, validating tool results before passing them to the model is a SHOULD, and so is the human confirmation step. Check whether your client implements them. The OWASP community page makes the same point about responses: the spec “advises clients to consider trust boundaries, but does not mandate response validation before passing content to the LLM.”
So the honest summary is that the protocol left these decisions to clients, and clients vary. Review the client implementation, server behavior and actual permission boundaries together.
Shrink the blast radius before you audit every tool
Description review is useful but cannot be the only control. Pair it with limits on what tools can access and transmit.
Review tool-description changes as security changes
Microsoft recommends giving metadata changes comparable review attention to system-prompt changes. That is a governance recommendation, not a claim that the model assigns both the same priority.
Gate sensitive disclosures and irreversible actions
The finance walkthrough ends with data leaving the building. Enforced permissions limit legitimate tool operations, but implementation vulnerabilities can break those boundaries. Also check what read-only access can disclose. Limiting AI agent permissions covers how to draw that line, and the short version is to gate sensitive data leaving the approved boundary as well as actions that cannot be undone.
Know what is connected before you try to review it
You cannot review a list you do not have, which is the problem shadow MCP servers creates. Start there if your answer to “which tools does our agent have?” is a guess.
Be deliberate about always-allow
Always-allow removes a human checkpoint for the covered calls. It does not disable every other safeguard or guarantee that a malicious action will succeed. Use it only for operations whose effects and data exposure are within an approved boundary; retain meaningful review for sensitive actions and inspect the actual arguments, not just the tool name.
FAQ
QCan a poisoned MCP tool affect me if I never call it?
Can a poisoned MCP tool affect me if I never call it?
QCan a scanner detect a poisoned tool description?
Can a scanner detect a poisoned tool description?
QDoes “always allow” mode make tool poisoning worse?
Does “always allow” mode make tool poisoning worse?
QHow do I inspect an MCP tool’s description?
How do I inspect an MCP tool’s description?
description field for each one, either through your client’s tool or connector panel if it exposes full descriptions, or by querying the server directly and reading the tools/list response. Inspect the full field where present. Connecting to an untrusted local server can itself execute code, so review its source first or inspect it in a restricted test environment without real credentials.Sources
- Microsoft Security Blog: Securing AI agents, when AI tools move from reading to acting
- MCPTox: A Benchmark for Tool Poisoning Attack on Real-World MCP Servers (arXiv 2508.14925)
- OWASP MCP Top 10 (MCP03:2025 Tool Poisoning)
- OWASP community attack page: MCP Tool Poisoning
- Model Context Protocol specification, Tools (2025-06-18)
Microsoft’s finance scenario is an attack pattern published by Microsoft Incident Response, not a disclosed breach at a named organization. MCPTox figures are from the August 2025 paper; the 72.8 percent result is the highest of the twenty agents tested, not an average. The OWASP MCP Top 10 was at beta stage when this was written.