A malicious MCP server was asked where to send the user to log in. It answered with something the operating system ran instead.
JFrog disclosed this in July 2025. It scored 9.6 out of 10 on the standard severity scale, which is about as bad as these get.
mcp-remote is the proxy that connects clients like Claude Desktop and Cursor to remote servers. On connect, it asks the server for its OAuth metadata, including the address to open in a browser for login.
A malicious server returns a value that is not an address. The proxy hands it to the operating system, and on Windows it executes. Command execution on the developer's machine, before a single tool ran.
Almost all MCP security work right now points at the tool layer. Meanwhile every client still takes an unknown server's word for how authentication is supposed to work.
Metadata from a server you do not trust is untrusted input.
JFrog's full write-up, including the technical chain:
champ.ly/yLoqeIf-
CVE-2025-6514, affecting mcp-remote 0.0.5 through 0.1.15. If you are on an affected version, update to 0.1.16 and connect only over HTTPS.
#MCP #OAuth #AppSec #AIAgents