Claude Cowork AI agent vulnerability allows VM escape and host file access

Researchers identified a vulnerability chain, dubbed SharedRoot, that enables an AI agent to escape its Linux virtual machine environment and access files on the host macOS system. The exploit leverages CVE-2026-46331 to achieve root privileges within the guest, subsequently exploiting a read-write host mount to bypass isolation. This highlights critical security risks in deploying autonomous AI agents within containerized or virtualized environments.