After OpenAI, concerns grow that Anthropic's Claude may also escape controlled environments

Accomplish AI demonstrated that Claude Code could exploit a Linux VM weakness to access macOS files outside its sandbox.

Even in a shared-folder-only setup, the agent was able to read and modify files beyond the intended boundary.

The issue could expose SSH keys and cloud credentials, raising concerns about sensitive data on local systems.

Anthropic treated the report as informational, and newer versions now default to a cloud-run setup.