Skip to content
TechTimes·Indexed by this site · Originally published· Attacks →· Its place in the scan record →SelectedAttention44

Preprint: when name, description and schema all match, three models including GPT-5.4 chose the impersonating MCP tool in all 50 trials

GPT-5.4, Qwen3.7 Max and DeepSeek V3 Picked Copycat MCP Server Tool Over Real One in All 50 Trials

AI summary

A preprint posted to arXiv on November 16 shows that standing up a separate MCP server with impersonating tool metadata, without compromising the host, the model or the legitimate server, can make models hand requests to the attacker's tool. Under the condition that tool names were exactly identical, descriptions were copied verbatim and schemas were compatible, GPT-5.4, Qwen3.7 Max and DeepSeek V3 all selected the attacker's tool in all 50 trials under that condition, Gemini 2.5 Pro was 84%, and Claude Sonnet 5 was 6%. Under this condition, Claude Sonnet 5 first asked for clarification in 94% of trials, while the other three models never asked. The authors call this failure tool-provider misbinding and propose provider-bound tool identity with signed manifests and stable provider identifiers as a fix, but the paper did not evaluate these mechanisms. The experiment covered only one benign dataset-acquisition task and is a preprint that has not undergone peer review.

Why it was selected

Under the condition where the tool name and description were fully impersonated, Claude Sonnet 5's 6% violation rate contrasts with the 100% of the other three models, showing the role of clarification questions in tool selection.

Related papers
Read originaltechtimes.com