GPT-5.4, Qwen3.7 Max and DeepSeek V3 Picked Copycat MCP Server Tool Over Real One in All 50 Trials
Reported by Tech Times1 source ↓
A preprint reports that three models selected an attacker-controlled copycat MCP tool in all 50 trials when its name, description and schema matched the legitimate tool. The attacker needed only to run a second server, not break into another system.
Key points
- The attacker controlled only a second MCP server advertising a competing tool
- Gemini 2.5 Pro selected the copycat in 84% of the matching-tool trials; Claude Sonnet 5 did so in 6%
- The study reports that an exact tool name alone rarely led models to choose the copycat
- Claude Sonnet 5 asked for clarification in 94% of the matching-tool trials
Details
Tech Times reports that the findings come from a preprint posted on arXiv on November 16. The authors call the issue “tool-provider misbinding”: an MCP client can present tool metadata from multiple servers, and a model may choose a tool based on that metadata without reliably identifying its provider. The MCP specification says tool names should be unique within a server, while clients combining servers may encounter name collisions and should use a disambiguation strategy.
In the key condition, the copycat used the legitimate tool’s exact name, copied its description and provided a compatible schema. GPT-5.4, Qwen3.7 Max and DeepSeek V3 chose it in all 50 trials each; Gemini 2.5 Pro did so in 84% and Claude Sonnet 5 in 6%. The paper reports that Claude asked for clarification in 94% of those trials, while the other three models never did.
The experiment tested five models across 12 combinations of tool name, description and schema. Ten prompts were repeated five times per model and condition, producing 50 trials per cell; the paper says every trial resulted in execution. An exact name without a convincing description and compatible schema had little effect in one tested condition: GPT-5.4 chose the attacker’s tool in 2% of trials, and the other four models did so in none.
Written from coverage as of