AI Agent Has Root

Running an unsandboxed MCP (Model Context Protocol) server gives an AI agent full access to the user's account permissions, including SSH keys, cloud credentials, and the ability to modify or delete files without any restrictions or audit trail. The author warns that even trusted MCP servers present significant security risks because trust is not static and the attack surface extends beyond just the server binary itself. Implementing proper sandboxing and security measures is critical before running MCP servers in AI agent workflows.