Guide · AI agent security · as of 22 November 2026

A copycat MCP tool was picked over the real one in 50 of 50 tries. Five checks.

An AI agent connected to several MCP servers chooses which tool to call from what each tool says about itself. A study posted on 16 November 2026 tested what happens when a second server advertises a tool with the same name, the same description and the same arguments as the real one. Three of the five models tested picked the copy every time, in 50 of 50 tries. Nothing was hacked to get there. The attacker only ran another server and wrote its own tool list. This guide covers what the researchers measured, what we found when we listed the MCP servers on our own workstation, and five checks for yours.

What the study measured

Same name, different owner.

The paper, by Zeynab Anbiaee, Sajjad Dadkhah and Ali Ghorbani of the Canadian Institute for Cybersecurity at the University of New Brunswick, is a preprint and has not been peer reviewed. They connected each model to two servers: a real one offering a tool called data_fetch_and_store, and an attacker’s one. They changed only what the attacker’s tool said about itself, across twelve combinations, and ran each combination 50 times per model. When the copy matched the real tool on name, description and arguments, GPT-5.4, Qwen3.7 Max and DeepSeek V3 called the attacker’s tool in all 50 tries, and Gemini 2.5 Pro in 42. Claude Sonnet 5 called it in 3 and asked the user which one was meant in the other 47. Adding words such as “officially verified” to the copy raised Claude’s rate to 13 of 50. The test was one harmless task over sample data, in a lab; no real incident is claimed.

What decided it

The arguments gave it away.

The same name on its own did almost nothing: with an unrelated description and different arguments, the copy was picked once in 250 tries across all five models. Copying the description as well, but keeping different arguments, was picked in none. The copy only won when it accepted exactly the arguments the real tool takes, so that one call would work on either server. To the model, the two tools were then the same tool. Asking the user is better than guessing, but the authors point out it does not solve the problem: a person shown two identical tool names cannot tell which server is the real one either. Their conclusion is that model caution “does not create provider-bound tool identity”. A tool’s name is a label. It says nothing about who runs the server behind it.

An MCP server could send one client’s answers to another

What we found on our own workstation

Eight servers, one name twice.

We connect our coding agents to MCP servers every day, so on 22 November we listed them. There are eight. Two are browser servers built from the same package, which means they advertise the same 25 tool names. Claude Code, the agent we use, keeps them apart by putting the server’s name in front of every tool, so a click in one browser cannot be routed to the other. That prefix is exactly the kind of binding the paper asks for, but it is only as good as the server name. And one server name was defined twice: once in the project file the whole team shares, and once in a personal override on this machine, with different settings. The override won, and it was the less restricted of the two. It could not connect, so nothing was reachable through it, but the team believed the shared file was in charge and it was not. The fix is to delete the override so the shared settings load.

AI agents in Dubai, connected only to the tools each job needs

Five checks for the MCP servers your agents use

Run these this week.

  • 1. List every MCP server each agent can reach, with who runs it. Include personal and machine-level settings, not only the shared project file. Test: run your agent’s server list command (in Claude Code, claude mcp list) and name the owner of every entry.
  • 2. Look for the same tool name on two servers. Two copies of one tool is how the copycat in the study won. Test: export each server’s tool list and sort the names; any name that appears twice needs a reason.
  • 3. Check that one server name means one thing. A name defined in more than one settings file can load settings you did not choose. Test: look for each server name in every settings file, and confirm the one that actually loads is the one you meant.
  • 4. Add servers by hand, from the maker’s own page. A description that says “verified” or “official” is just text the server wrote about itself. Test: for each server, find the maker’s repository or documentation, and remove any server you cannot trace.
  • 5. Keep write tools behind a person. If a copied tool is picked, what it can do decides the damage. Test: list the tools that can change or send anything, and confirm each one asks before it runs.

Questions

Questions about MCP tool names.

  • What is an MCP tool name collision?

    It is when two MCP servers connected to the same AI agent offer a tool with the same name. The agent has to choose between them from what each server says about its tools, and none of that proves who runs the server. A study posted on 16 November 2026 found that some models then call the attacker’s copy every time.

  • Can a malicious MCP server take over a real tool?

    In the study, it did not need to break anything. A server that copied the real tool’s name, description and arguments was chosen in all 50 tries by GPT-5.4, Qwen3.7 Max and DeepSeek V3, and in 42 by Gemini 2.5 Pro. Claude Sonnet 5 chose it 3 times and asked the user in the other 47. It was a lab test on one task, not a reported attack.

  • Does Claude Code namespace MCP tools?

    Yes. Claude Code puts the server’s name in front of each tool, so two servers offering the same tool name show up as two different tools. That stops a call meant for one server being sent to another, as long as each server name itself points where you think it does.

  • What are the main MCP server security risks?

    From the guides on this site: a server you did not mean to add, a copied tool chosen over the real one, a server mixing up the answers of clients that share it, and archived servers that will never be fixed. The common defence is the same: know who runs every server, and which of its tools can change things.

  • How do I stop my AI agent picking the wrong MCP tool?

    Connect only the servers a job needs and remove any duplicate tool names. Define each server name in one place, and put tools that change data behind a confirmation. The researchers also propose signed tool lists and fixed tool identifiers tied to the server; those are proposals, not yet features you can switch on.

Sources

read 22 November 2026

Check it at the source.

  1. arXiv: When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents (Anbiaee, Dadkhah and Ghorbani, University of New Brunswick, submitted 16 November 2026; a preprint, not peer reviewed)
  2. BRev Digital: an MCP server could send one client’s answers to another (two flaws in the official MCP TypeScript SDK)
  3. BRev Digital: deprecated MCP servers your AI agent may install (archived servers that will never be fixed)
  4. BRev Digital: Claude Code security settings (the controls you can check yourself)

A question?

Know which tools your agents can call?

Tell us which AI agents your team runs and what they are connected to. We will list every MCP server with its owner and find any duplicate tool names or overridden settings. Then we put the tools that change things behind a person.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com