Security1 hOctober 12, 2026, 19:07

‘Repeat-After-Me’ prompt bypasses six AI agent frameworks

Researchers from Penn State, Université Montpellier and Dartmouth tricked six widely used AI agent frameworks into executing attacker-written tool calls by simply asking the model to repeat the text. The technique worked on LangGraph, Microsoft Agent Framework, AutoGen, Google ADK 2.0, AWS Strands Agents and CrewAI. Their ToFu fuzzer confirmed 253 bugs across 118 of 273 open-source MCP servers.

CSO Online

Open source

‘Repeat-After-Me’ prompt bypasses six AI agent frameworks

Attack worked on LangGraph, AutoGen, ADK 2.0, Strands Agents and CrewAI
The model refused, but the framework ran the call before the refusal
ToFu found 253 bugs in 118 of 273 MCP servers
Bugs include 64 path traversals, 7 SSRF and 3 OS command injections

#LangChain#Microsoft#Google#AWS

LLM-RedKit: agentic prompt injection yields 10 findings in 16 attempts

Resilience test: how six agent frameworks recover from eight broken tool calls

Policy

US Accuses Six Chinese AI Firms of Copying Models, Bypassing Chip Controls

Security

‘Repeat-After-Me’ prompt bypasses six AI agent frameworks

LLM-RedKit: agentic prompt injection yields 10 findings in 16 attempts

Resilience test: how six agent frameworks recover from eight broken tool calls

US Accuses Six Chinese AI Firms of Copying Models, Bypassing Chip Controls

CVE-2026-61439: Prompt injection defense bypass in PraisonAI