7 Sandbox Escape Vulnerabilities Across 4 Coding Agent Vendors

4 points by afogel


emk

I feel very strongly that:

  1. Anything based on asking users to approve/deny is a catastrophe waiting to happen.
  2. Allowing some other model to decide whether to approve or deny (like Claude Code's auto mode) is even sketchier.
  3. OS-level sandboxing is great, but not if you allow agents to have read-only access $HOME/Documents/Taxes/. You need to hide anything outside the working directory, and maybe some approved dotfiles.
  4. Agents shouldn't be trusted with real credentials. Sorry. Give 'em an HTTP proxy that injects appropriately limited API keys where needed.
  5. Web-browsing MCPs should convert output to Markdown and ideally try to filter out prompt injections. Not that they'll succeed, but even trying reduces the exposure.

Unfortunately, while Linux supports this kind of sandboxing, actually setting it up requires some knowledge to configure actual, minimal permissions. And of course, a truly malicious and clever agent could find a kernel vulnerability and escape. But that's a more drastic threat model.

So these recommendations are not technically feasible in most environments, especially outside of custom Linux setups.