Claude Cowork's "SharedRoot" flaw let AI agents escape their sandbox and read Mac files
Security researchers at Accomplish AI showed that a single message to a local Claude Cowork session could break out of its Linux VM via a kernel exploit, reaching the host Mac's filesystem — including SSH keys and cloud credentials — with no permission prompt. Anthropic closed the report as informative; Cowork now defaults to cloud execution, which sidesteps the local escape path.
why it mattersA one-message sandbox escape on roughly 500,000 local Cowork installs is a stark reminder that agentic coding tools need real isolation, not just a VM label.