Sandbox-CLI is now in public beta (sandbox-cli.vercel.app)

Sandbox-CLI has launched in public beta, allowing developers to run various coding agents like Claude Code, Codex, and Gemini within isolated Docker containers. This innovative tool emphasizes security by providing agents with full autonomy while preventing access to sensitive host files such as SSH keys and cloud credentials, restricting interactivity to a specified project directory. It uses ephemeral paths that disappear after execution, ensuring that no extraneous data is exposed. This release is significant for the AI/ML community as it allows for safer experimentation with coding agents by limiting their potential blast radius to the specified workspace. Key features include strict default-deny rules for host environment variables, mandatory path specifications, and the ability to opt-in for additional controls like credential management and egress allowlists. These measures mitigate the risks of prompt injections that could otherwise lead to unauthorized code execution. The combination of enhanced isolation protocols and straightforward Docker integration makes Sandbox-CLI a promising tool for developers looking to leverage AI/ML coding agents securely and efficiently.