Researchers bypass 6 AI agent frameworks with a simple ‘Repeat-After-Me’ prompt

Summary

Researchers from Pennsylvania State University, Université Montpellier and Dartmouth College showed that a "Repeat-After-Me" prompt can make six AI agent frameworks, including LangGraph, Microsoft Agent Framework, AutoGen, Google's Agent Development Kit 2.0, AWS Strands Agents and CrewAI, run attacker-written tool calls. In a LangGraph 1.2.0 demonstration with Llama 3.1 8B, the tool executed before the model's refusal arrived. Their ToFu fuzzer also confirmed 253 bugs affecting 343 tools across 118 of 273 open-source MCP servers.

Topics

• AI agents
• Model Context Protocol

Related items

• InfoMeta debuts Muse Wonder AI model in latest test of users' willingness to paySame vendor · CNBC Technology
• InfoMeta has some Wonder Full new AI models Same vendor · The Verge (AI)
• InfoIs AI Coding Safe? Introducing the Agent Security League | Blog | Endor LabsSame vendor · Endor Labs Blog
• InfoMicrosoft’s new AI chip makes it less reliant on Nvidia’s GPUsSame vendor · The Verge (AI)
• InfoBring Your Own Agent: Your Employees' Assistants Are Already Logged InSame vendor · The Hacker News