Large language models are increasingly used as intermediaries in Model Context Protocol (MCP) systems, where they decide when and how to call external tools. A new paper on arXiv introduces COPEX, a benchmark designed to measure how robust those models are to adversarial context. The authors identify four entry points for adversarial influence: user instructions, tool schemas, tool outputs, and protocol messages.
Because the abstract is truncated, the paper's exact comparison with prior benchmarks is not fully available. What is clear is that COPEX is positioned as a response to existing benchmarks, which the authors argue do not adequately cover this setting. The benchmark appears to treat each MCP layer as a separate attack surface, rather than treating adversarial input as a single generic problem.
The significance is practical: as LLMs take on more tool-use responsibilities, attackers may target the context around the model rather than the model itself. A benchmark like COPEX could help developers see where their systems are most exposed before deployment.