Claude Cowork flaw lets AI agent escape VM and access Mac files

Accomplish AI researchers found a sandbox escape vulnerability in Anthropic's Claude Cowork that lets an attacker break out of the agent's Linux VM to read or write files anywhere on the Mac.