Полезайте в песочницу, мистер Claude: изолируем агента

The article describes using Docker Sandbox (CLI: sbx) to run AI agents such as Claude Code inside a microVM, so the agent cannot directly access host secrets like OAuth tokens, SSH keys, or GitHub PATs. It highlights key isolation properties—separate microVM kernel, no access to host Docker daemon/socket, workspace-only filesystem exposure, and network restrictions that force outbound traffic through a host-side proxy—making accidental destructive actions or token theft less likely. No specific CVE identifiers are mentioned, but the core message is that stronger isolation at runtime matters when agents run with the permissions of the user who launched them.