AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

MCP-SafetyBench: A Benchmark for Safety Evaluation of Large Language Models with Real-World MCP Servers

Xuanjun Zong, Zhiqi Shen, Lei Wang, Yunshi Lan, Chao Yang

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 25 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
prepend a safety prompt to the user requestall models · MCP-SafetyBenchweighted ASR (%): 39.88 -> 38.65No clear effect
safety promptall models · Malicious Code Execution attacksASR: -21.54%Improved
safety promptall models · Credential Theft attacksASR: -21.37%Improved
safety promptall models · Function Overlapping attacksASR: +9.36%Hurt
safety promptall models · Preference Manipulation attacksASR: +7.34%Hurt
safety prompto4-mini · MCP-SafetyBenchASR: -8.98%Improved
safety promptopen-source models (Kimi-K2) · MCP-SafetyBenchASR: +5.30%Hurt
add more MCP servers (2 to 4) (ablation)GPT-4.1 · financial domainTSR (%): 24.53 -> 30.19Improved
add more MCP servers (2 to 4) (ablation)GPT-4.1 · financial domainASR (%): 43.40 -> 47.17Hurt
use reasoning vs non-reasoning modelLLM agents · MCP-SafetyBenchASRNo clear effect
use open-source vs proprietary modelLLM agents · MCP-SafetyBenchASRNo clear effect
higher task success rateLLM agents · MCP-SafetyBenchdefense success rate: r = -0.572Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy