Docker Desktop 4.63 introduces 'docker agent' with optional sandboxing

Docker Desktop version 4.63 has quietly introduced a new command-line tool, 'docker agent,' designed to build and share AI agents using declarative YAML configurations. The tool integrates with various AI models and MCP (Model Context Protocol) servers. While the feature offers a robust security sandbox, including a default-deny egress proxy and restricted filesystem access, these safety measures are disabled by default. Security researchers note that running agents without the '--sandbox' flag exposes the host machine to significant risks, as the agent operates with the user's full permissions. The documentation emphasizes that while the architecture is well-designed for isolation, users must manually opt-in to these protections. The release highlights the ongoing tension between developer convenience and secure-by-default practices in the rapidly evolving AI agent ecosystem, prompting calls for Docker to reconsider its default security posture in future updates.