Description

A critical security flaw, identified as CVE-2025-49596 with a CVSS score of 9.4, was discovered in Anthropic’s Model Context Protocol (MCP) Inspector tool. This vulnerability enables remote code execution (RCE), allowing attackers to fully control the host system. It represents one of the first major RCE threats within Anthropic’s MCP ecosystem, exposing a new class of browser-based attacks targeting AI development tools and posing serious risks to developers, enterprises, and open-source projects. The MCP Inspector tool is built to facilitate the testing and troubleshooting of MCP servers, which enable AI models to connect with and utilize external data sources. It functions using both a client-side interface and a proxy component working together. However, its default configuration is insecure, missing critical protections like authentication and data encryption. This oversight can leave the proxy exposed to unauthorized access, especially from untrusted networks, allowing malicious actors to trigger local processes or implant backdoors through direct interaction with the server. The exploit merges an outdated browser vulnerability called "0.0.0.0 Day" with a CSRF weakness present in the MCP Inspector tool. Malicious websites can send unauthorized requests to the proxy running on localhost (127.0.0.1), resulting in arbitrary code execution. Techniques like DNS rebinding can also redirect traffic to local services, bypassing browser security and giving attackers control over the machine. In response to the responsible disclosure in April 2025, Anthropic resolved the security flaw by issuing an updated release—version 0.14.1 of MCP Inspector—on June 13. The patch implements session tokens and adds origin validation to effectively block the exploitation route. Researchers warned that localhost services, though seemingly secure, are often vulnerable due to browser behaviors and insecure defaults—urging developers to apply the fix and harden their configurations.