AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

SOPE: Situation-Aware and Statistically Indistinguishable Privacy Exfiltration for MCP-enabled Agents

Ruixiao Lin, Qingming Li, Jiahao Chen, Chunyi Zhou, Shouling Ji

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agent safety and attack benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SOPE, a situation-aware, statistically indistinguishable privacy exfiltration attack against MCP-enabled LLM agents that exploits malicious tool supply chains.

Problem

Privacy leakage in MCP-enabled LLM agents via indirect prompt injection; existing attacks are misaligned with the agent's tool-usage context and are easily detectable.

Approach

Crafts exfiltration payloads that are aligned with the agent's situational tool-usage context and statistically indistinguishable from benign behavior, enabling covert privacy exfiltration through MCP tool supply chains.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add the trained Assign module (ablation)SOPE · Evasion against retrieval screeningESR: 0.44 -> 0.87Improved
add the trained Assign module (ablation)SOPE · Targeted, benchmark agentsleakage rate: 0.78 -> 0.92Improved
add the trained Assign module (ablation)SOPE · Targeted, benchmark agentsASR (total): 0.77 -> 0.75Hurt
replace frontier-LLM binary query with trained Assign (ablation)SOPE · Targeted, benchmark agentsleakage rate: 0.62 -> 0.92Improved
add optimization-based Fuse instead of appending probing prompts (ablation)SOPE · Targeted, benchmark agentsASR (total): 0.23 -> 0.75Improved
add optimization-based Fuse instead of LLM rewriting (ablation)SOPE · Targeted, benchmark agentsASR (total): 0.41 -> 0.75Improved
add code-level Leak (transmit) instead of explicit instruction (ablation)SOPE · Targeted, benchmark agentsASR(S2): 0.46 -> 0.86Improved
add code-level Leak (transmit) instead of explicit instruction (ablation)SOPE · Targeted, benchmark agentsASR (total): 0.39 -> 0.75Improved
deploy IsolateGPT system-isolation defense against SOPESOPE · Targeted, benchmark agentsleakage rate: 0.92 -> 0.83Hurt
deploy retrieval-screening defenses against SOPESOPE · Targeted, benchmark agentsASR (total): 0.75 -> 0.88Improved
target commercial agents instead of benchmark agentsSOPE · Targeted attackleakage rate: 0.92 -> 0.94Improved
use SOPE instead of indirect-prompt-injection baselinesbaseline attacks · Commercial agents, averageASR: 0.35 -> 0.96Improved

Experimental setup

Builds on: Model Context Protocol

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