Tiktok 论文发现：MCP 工具描述可诱使 AI 助手完成任务后泄露对话Chinese only
研究人员将诱导指令藏进 MCP 工具描述，让 AI 助手在完成正常请求后调用攻击者控制的日志工具，记录用户提问、服务器回复和最终答案。在可运行工具的 18 组模型与服务器组合中，论文报告的攻击成功率为 62.64%—100%。
1 developments2 reports1 sources
Event developments 1
Tiktok paper finds MCP tool descriptions can prompt AI agents to leak conversations after completing a task
Researchers hid instructions in an MCP tool description to prompt agents to call an attacker-controlled logging tool after completing the legitimate request, recording the user query, server response and final answer. Across 18 model-server pairs capable of running tools, the paper reports attack success rates of 62.64%–100%.
View source reports · 1 sources · 2 reports
- Rohan Paul (@rohanpaul_ai) · Tiktok paper finds MCP tool descriptions can prompt AI agents to leak conversations after completing a task
- Rohan Paul (@rohanpaul_ai) · Log-To-Leak study shows how MCP tool descriptions can induce AI agents to leak interaction records
All reports 2
Log-To-Leak study shows how MCP tool descriptions can induce AI agents to leak interaction records
Researchers placed an instruction in an MCP tool description to induce agents to call an attacker-controlled save_logs tool after completing the original task, sending the user query, server reply and final answer. Across 18 model-server pairs able to run tools, the reported attack success rate was 62.64–100%, with conditional logging accuracy of 76.55–95.64%.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)
Tiktok paper finds MCP tool descriptions can prompt AI agents to leak conversations after completing a task
Researchers hid instructions in an MCP tool description to prompt agents to call an attacker-controlled logging tool after completing the legitimate request, recording the user query, server response and final answer. Across 18 model-server pairs capable of running tools, the paper reports attack success rates of 62.64%–100%.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)