deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

Meta's Muse agent runs in a Secure VM with Sentinel approval for internet actions

Meta's Muse personal agent runs on a dedicated Secure VM with a Sentinel agent approving internet-facing actions and a user-visible audit trail, but the user-held-key Confidential VM has not shipped yet.

Meta's Muse agent runs in a Secure VM with Sentinel approval for internet actions

Meta has shipped a consumer AI agent with an uncommon security architecture. According to an analysis by security firm Netics published on dev.to, the Muse personal agent — introduced through Meta's newsroom on September 8 — runs on a dedicated virtual machine in the cloud, with a separate Sentinel agent approving every internet-facing action and a complete audit trail visible to the user.

On the surface, Muse is a familiar product: it works across the apps people already use, sends email, books travel and pays through Link. What Netics finds notable is the framing. Where most consumer agents are pitched as chatbots with permission settings, Muse is pitched as a sandboxed computer with enforced boundaries — a real operating envelope rather than a promise baked into the model's instructions.

Isolation, mediation and audit

The design rests on three properties. Isolation: Muse runs on its own dedicated cloud machine, contained so that no other agent can reach it, and that machine holds the user's data and the credentials for connected services. Mediation: nothing Muse does reaches the internet unless the Sentinel approves it, and the Sentinel asks the person for permission when needed. Audit: Muse displays a full record of everything it has done and everything it plans to do.

Credentials the agent cannot read

The credential handling is singled out as the strongest pattern. Muse reportedly has no visibility into passwords or payment methods; credentials a person shares go into secure storage so the agent can act with them without ever seeing them, including passwords typed into the browser. Netics likens this to enterprise practice in which a service is granted narrow, scoped access rather than the underlying secrets. The agent receives a capability to act, not the key itself, and laypeople get the benefit without understanding the architecture.

The user keeps the controls

The announcement hands users the controls a security engineer would ask for. People choose which apps Muse connects to and how much access each one gets — for email, whether Muse only reads mail or can also send on their behalf. Access can be changed or revoked at any time. Users can opt out of their interactions being used to train Meta's models and can tell Muse to forget specific things it has learned. Meta states that Muse does not share a person's conversations or VM data with its advertising systems.

Netics argues these map onto what it would specify for a business agent: least privilege by default, per-capability scope choices, revocation and an audit record. The consumer version is simpler and less configurable, but the direction is right, and the pattern transfers to any team building customer-facing agents.

What the announcement leaves open

The strongest privacy claim is not live yet. The Confidential VM, in which the entire machine — data and conversations included — is encrypted with a key only the user holds, so not even Meta can access it, is planned for later this year, according to the announcement. Until it ships, the trust model rests on Meta's own enforcement of the Secure VM boundary.

The Sentinel also raises unanswered questions. Its policy scope is not published, so what counts as sensitive, whether the user can override it, and how third-party connections are audited all remain open. The announcement likewise says little about the model side — what Muse learns, where it is stored and what happens on opt-out — precisely the details Netics expects European stakeholders to probe under the GDPR and the EU AI Act.

Why it matters

Agents that send email, book travel and move money are no longer just chat interfaces, and their safety can no longer rest on model behaviour alone. In Netics' reading, Muse is the first mainstream effort to put a consumer agent inside a genuine security perimeter: an isolated VM, mediated network egress, scoped credentials the agent cannot read, revocable access and a user-visible audit log. That blueprint applies directly to enterprise agents and is likely to become the benchmark other vendors are measured against. The caveat matters just as much: the claims currently run ahead of the evidence. Until the user-held-key Confidential VM ships, users are trusting Meta to police its own boundary, and the Sentinel's rules and the model-side data flows remain undocumented.

  • #meta
  • #ai-agents
  • #security
  • #privacy
  • #virtualization

Related posts