evaluation-package · Source-linked discovery

MCPTox

MCPTox measures whether tool-using LLM agents are manipulated by Tool Poisoning Attacks, where a malicious instruction is hidden inside a tool's description (the metadata an agent reads when planning) rather than in any executed code. Built on 45 live MCP servers and 353 authentic tools, it presents an agent with a benign user query and a server whose tool set contains one poisoned tool, and measures the Attack Success Rate via a model judge.

Open interactive record →
OriginZhiqiang Wang, Yichao Gao, Yanting Wang et al.TopicsSafeguardsStatuscatalogued

Can support

Not independently assessed by FronteraEval yet.

Cannot support by itself

No inference beyond the upstream source should be made until the protocol is reviewed.

Original sources