Saved report copy
Recorded
MCP trust gaps let attackers pivot between AI agents
Independent researcher Syed Anas Mohiuddin demonstrated attacks that exploit trust gaps in Model Context Protocol to get one AI agent to forward malicious instructions to another trusted agent. His testing involving systems from Google, JP Morgan Chase, Weviate, Rapid7, the French government's interministerial digital directorate and the US federal government found server-side request forgery issues, including a Google MCP database toolbox flaw rated 8 out of 10 and Rapid7 flaw CVE-2026-97228 rated 2.7 and fixed last month. Researchers said the findings matter because internal agents often trust each other by default and lack consistent authorization, making multi-step injection hard to catch.