Description

Security researchers from OX Security have identified a critical architectural weakness in Anthropic’s Model Context Protocol (MCP), a widely used standard for AI agent communication. This flaw enables arbitrary command execution, allowing attackers to take full control of affected systems. Exploitation could expose sensitive assets such as internal databases, API keys, chat histories, and user data. Because the issue originates from the protocol design rather than a coding mistake, it impacts multiple official MCP SDKs across languages like Python, Rust, Java, and TypeScript. The vulnerability has a large supply chain impact, affecting millions of deployments and thousands of publicly accessible servers. Developers using MCP-based tools may unknowingly introduce this risk into their environments. Attackers can exploit the flaw through multiple techniques, including unauthenticated UI injections, zero-click prompt injections in AI IDEs, and malicious package distribution. Several AI platforms, such as Flowise and others, have been exposed, leading to multiple CVEs across widely used tools and frameworks. Despite responsible disclosures, Anthropic has not modified the protocol, stating the behavior is expected. Organizations must adopt strong mitigations, including restricting public access, treating all MCP inputs as untrusted, and deploying services in isolated environments. Continuous monitoring, secure installation practices, and timely updates are essential to reduce the risk until a comprehensive fix is introduced.