Anthropic's MCP Protocol Design Flaw Affects Over 200,000 AI Servers, Impacting Tools Like Cursor

OX Security, a cybersecurity firm, revealed a design flaw in Anthropic's MCP (Model Context Protocol) that allows remote code execution, affecting over 200,000 AI servers. The flaw lies in the SDK's STDIO interface, enabling arbitrary OS command execution without validation or warnings.

Tools like Cursor, Claude Code, and Gemini-CLI are impacted. Anthropic stated the issue was 'expected behavior' upon notification, only updating documentation instead of fixing the architecture. Some tools like LiteLLM have released patches, while others like LangFlow are still pending fixes.