AI & Enterprise
Anthropic\'s Claude Cowork also raises fears of escaping controlled environment after OpenAI case
A similar weakness has been found in Anthropic\'s Claude Cowork after confirmation that an experimental OpenAI agent could escape a controlled sandbox. Security firm Accomplish AI demonstrated a Claude Cowork AI agent exploiting a Linux flaw to break out of a virtual machine on a Mac and access macOS files beyond an allowed shared folder. It estimated up to 500,000 local macOS users could have been exposed before an update.
