Meta Muse security explained: isolated cloud VM plus Sentinel review
The author breaks down Muse's security design: each user gets a private cloud Linux VM, Sentinel reviews actions externally against permissions, OAuth credentials are stored separately, and the agent only receives a stand-in token.
🔐 Breakdown of how Meta Muse security works 2 of the people building Muse at Meta (@dps and @bigT_sheesh) recently shared how it works under the hood. Here’s what you need to know: 💡 TLDR: your muse runs in its own cloud computer, and the part that can act is kept separate from the part that holds your passwords. 🔹 Every user gets a Muse Secure VM, a private computer in the cloud with its own Linux environment and browser. You can open the Library tab and see every file it’s working with. 🔹 Sentinel is a security agent that sits outside that environment. It checks anything leaving the VM against the permissions you set: allowed, denied, or ask you first. 🔹 When you connect an app like Gmail, the OAuth token (the key that lets Muse act on your account) is stored in a separate credentials service, outside the agent’s environment. Passwords go into a special screen in the app, never through the chat. 🔹 The agent only ever gets a stand-in token. Sentinel swaps in the real credential after the request is approved. Hope this is helpful! Been using @muse from @meta since launch & will share more use cases as info here. Have you been using personal agents?