3 AI models selected a copycat MCP tool in 50 of 50 trials
Claude Sonnet 5 selected the copycat in only 6 percent of trials. It usually gave the user a question.
On 16 November, researchers put a preprint on arXiv, a paper site. It reports that 3 AI models selected a copycat tool on a different MCP server, and not the correct tool, in all 50 trials. The attacker did not get access to the host, the model or the correct server, and only operated a server with set tool data.
How it works
The MCP client gets the tools of each connected server. Each tool has a name, a description and an input format. The model reads this data and selects one tool. MCP is the Model Context Protocol, a standard that connects AI models to external tools. The protocol does not make sure that each server name is different.
The numbers
The attacker used the correct name, the description of the correct tool and a compatible input format. The copycat was then selected in these rates of trials:
- GPT-5.4, Qwen3.7 Max and DeepSeek V3: 100 percent each
- Gemini 2.5 Pro: 84 percent
- Claude Sonnet 5: 6 percent
What is not known
The authors give a solution: signed tool data that names the provider. They did not test it. A question to the user does not show which provider is behind each tool. The paper reports no test with users.
Insight: Trust in an MCP server comes only from what it says about itself
In November, a bug in the MCP kits could send login secrets to a bad server. An October paper found bugs in 118 of 273 servers. Then a copycat server gets selected with the same description. The attacker only operates a server, and MCP cannot show which provider made a tool. The next move to monitor can be a signed identity standard. This adds risk with each added server.
Sources
Posted