How We Built Safety Into Muse — Muse
Summary
Muse secures its AI agent by using an isolated harness and a non-overridable Sentinel to restrict network egress and third-party service actions.
Isolated VMs protect user data.
Muse assigns a dedicated Linux box to every user to store credentials and run tasks like compiling code or managing sub-agents.
Each Virtual Machine(VM) is an isolated linux box
Sentinel controls all external actions.
As a separate host-side agent, Sentinel acts as the sole authority for network egress and third-party service connector permissions.
Sentinel is a separate host-side agent from your Muse
Surrogate tokens prevent credential leakage.
Sentinel performs just-in-time insertion, replacing surrogate tokens with real credentials at the network boundary so the agent never sees secrets.
Single-use cards secure financial transactions.
To prevent theft via prompt injection, Muse issues single-use card numbers and requires human-in-the-loop approval for every purchase event.
a single-use card number is issued
Confidential VMs aim to block Meta.
The upcoming Muse Confidential VM plans to use cryptography to verifiably prevent Meta from accessing any data within the user's VM.
cryptographically and verifiably prevent Meta from accessing data
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.