vmforai runs agents that act without asking, gives phones a way into your VMs, and keeps tokens in a few files. If you find a way around any of that (the phone page's key, the restricted network, the files' permissions, or anything else), please report it privately: Report a vulnerability.
You will get an answer within a few days, and credit in the release notes if you want it.
What vmforai does and does not promise:
- A VM keeps the agent away from your computer's files and keys, as far as the hypervisor does. It is not a defence against a determined attacker who breaks out of QEMU/KVM.
- The agent has the VM's network unless you restrict it (
vmforai net NAME restricted), and the credentials you gave it there (Claude's login, an API key) can be read by what runs in the VM. - The phone page is as safe as its link: the key in it gives access to your
VMs' terminals. It listens on Tailscale (encrypted) or, with
--lan, on your local network without encryption.