Meta runs Muse agents in CPU-only AMD sandboxes
An offer of SSH access to Muse’s private VM raises questions about what the agent could do, though harmful commands had not occurred.
Meta runs its Muse agent in CPU-only sandboxes on AMD EPYC Turin systems, with two cores and 8GB of memory allocated to each sandbox. Evan Hoffman and Tae Kim found that the virtual machines report AMD EPYC 9D25 processors, while Meta uses separate GPU servers for inference. Muse had reportedly exceeded 500,000 daily active users a few days before the report. Hoffman said Muse offered to configure SSH access to its private VM; the article describes a possible route to more harmful commands, but says that has not happened.
Why it matters
For users, the issue is the scope of an agent’s access, not just its answers: SSH offers a route into a private virtual machine. The report describes more damaging commands as a possibility, not an event, so it does not establish that Muse acted harmfully. Its reported reach makes that distinction relevant beyond a limited trial.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.