Skip to content
by InfinyAI ES

IA en un minutoNewsSecurity

SecurityUpdated:

A researcher finds the same security flaw in AI agents from Google, JPMorgan and the French government

IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds

Google, JPMorgan Chase, Weaviate, the French DINUM and the city of Tangerang in Indonesia all had the same hole in their MCP servers, according to Syed Anas Mohiuddin. All it took was a hidden order inside some content: the first agent passed it on and the second obeyed, because it trusts the first. Each organization confirmed it and has already patched it.

Independent researcher Syed Anas Mohiuddin has found the same kind of flaw in MCP servers run by organizations that have nothing in common: Google, the bank JPMorgan Chase, the database company Weaviate, the French government's interministerial digital directorate (DINUM) and the city government of Tangerang, in Indonesia. MCP (Model Context Protocol) is the open standard that AI agents use to connect to tools and data, and MCP servers are the connectors that give them those tools. Ars Technica reported the findings.

The problem is trust. One agent reads content with a hidden instruction, passes it to another agent as a normal task, and the second agent carries it out because it trusts the first one. In many cases, the result is a server making requests to the internal network on the attacker's behalf. Mohiuddin calls it "protocol pivoting"; Markus Vervier, of the security firm X41 D-Sec, sees it as a form of indirect prompt injection.

Each organization confirmed the flaw and has fixed it. Google's, in its MCP toolbox for databases, was rated 8 out of 10 for severity (CVE-2026-14540). The cybersecurity company Rapid7 published another related flaw, rated low. Mohiuddin has also reported issues in five MCP servers built for the US federal government, which are still under review and which he does not present as confirmed.

Douglas McKee, of Rapid7, summed it up for Ars Technica: "anything passed from an LLM to your tool should be treated like input from a stranger on the internet."

Why it matters · analysis and opinion

That Google, a major bank and a European government stumbled over the same mistake shows that the problem is not a bad programmer but a habit: taking for granted any order that comes from another agent. That chain of trust is what lets an attacker hide an instruction in a piece of content and end up making requests inside the internal network. Every organization has fixed it, but any company that connects agents to its databases may have the same gap without knowing it. Rapid7's advice serves as a working rule: treat whatever comes out of a model as if it came from a stranger. In practice, that means an inventory of the installed MCP servers, minimum permissions for each connector and up-to-date patches.

Source: Ars Technica · Written with the help of AI: how we make the news

AI in your inbox, every day or every Friday

The stories that matter, each one in a minute. With the source for every one.

Choose one or both:

Sign up and you are in: the daily arrives every night and the weekly on Friday mornings. You can unsubscribe from any email.