News

Sandbox Escape in Claude Cowork

Researchers have discovered a vulnerability allowing Anthropic's AI agent to escape a Linux VM onto the host macOS.

Compiled by Sergey KostenchukPublished 2026-07-25Updated 2026-07-25

2026-07-25 Safety Anthropic

Expanded analysis for this story

Open the longform version with context, source trail, and what changed.

Read longform

Escaping Claude Cowork's local VM sandbox via CVE-2026-46331 Source

🛡 Sandbox Escape in Claude Cowork

Researchers have discovered a vulnerability that allows Anthropic's AI agent to escape a Linux VM onto a host macOS via CVE-2026-46331. This provides access to SSH keys and user cloud data.

🌍 The industry needs to move toward host-side isolation.

👤 AI agent users should use scoped mounts to protect their data.

Sources

accomplish.ai