共有:

Meta Muse Security

The design idea:
never hand the agent
a real key.

How do you safely let an AI agent reach into your email, calendar and payments? Meta's answer with Muse is not to teach the model to behave — it is to cut off what the agent can touch, structurally. A look at the one-VM-per-user design and at Sentinel, the separate process that owns every outbound connection.

AI Navigate Editorial2026.09.107 min read

MUSE SECURE VM Agent Browser Files / Memory It only ever sees placeholder tokens SENTINEL Injects the real secret External services
FIG. The agent stays inside its isolated VM; every outbound request is judged by Sentinel
01
The Architecture

From "teach the model"
to "cut it off structurally"

On September 8, 2026, Meta launched Muse, a personal AI agent, for adults in the United States across iOS, Android and the web — with access through WhatsApp as well. Internally it had been tested under the codename "Hatch," with employees letting the agent operate third-party apps and browse the web on their behalf. Meta published its design rationale in an engineering post, How We Built Safety Into Muse.

At the center is the Muse Secure VM: a dedicated cloud machine created for each individual user. The agent itself, its data, its browser, its files and the credentials for connected services all live inside that one box. The premise is that when something goes wrong — a mistake or an attack — the blast radius stops at a single machine.

The sharper idea is Sentinel. It is a separate process, isolated from Muse at the system level, and in Meta's words it is "the sole permission authority for approval to perform actions with connectors to third-party services and for all egress over the network." The agent only ever sees placeholder tokens; Sentinel injects the real secrets at the network boundary. So a prompt injection attempting to exfiltrate credentials finds that there is nothing real inside the agent's field of view to steal.

02

The terms at launch

100M
free tokens per week
$20 / $100
paid plans (per month)
09.08
launch date (2026, US)

Where an agent runs shows up directly in its pricing. Standing up one VM per user costs more per unit of compute than piling everyone onto a single shared inference fleet the way a chatbot does. Publishing the free allowance as a token count is the flip side of that weight.

Muse is free up to 100 million tokens a week, with $20 and $100 monthly plans for people who need more compute. Meta has also signalled a future "Confidential VM" variant, where each VM runs inside a trusted execution environment and the user holds the keys.

Conventional (leave it to the model)Muse (separate it structurally)
Assumes the model refuses dangerous instructionsEven if it fails to refuse, it cannot reach the keys
Credentials sit inside the agent's contextThe agent only ever sees placeholder tokens
Execution and permission live in one processSentinel judges egress from a separate process
Damage can spread across every connected serviceDamage is meant to stop at that user's own VM

The best way to stop a theft
was to leave nothing worth stealing.


03
Why It Matters

How to read this design

01

Why now

Agent safety has long centered on getting the model to refuse dangerous instructions. But once an agent can reach email and payments, one bad judgment turns into real damage. Muse's answer is separation of authority — the AI that decides and the gatekeeper that holds the keys are different things. That is classical security design imported into agents, and it moves the industry conversation from "how smart is the model" to "what happens when it breaks."

02

Who it helps, and how

For consumers, a malicious email or web page that hijacks the agent still struggles to walk away with credentials. For developers building their own agents, the pattern is directly copyable: show the executing agent fake tokens, inject the real ones at the egress boundary. For IT teams evaluating this for work, the caveat is that Muse is a consumer product — do not assume audit logs or SSO controls.

03

What to do next

Start with low-stakes accounts, and do not blanket-approve what Sentinel asks you. The approval prompt is the last line of judgment; skim it and the architectural benefit evaporates. If internal company data might ever be in scope, waiting for the user-held-key Confidential VM version is a defensible call.

04

The counterpoint

A sound design is not the same as a sound implementation. Forbes and implicator.ai report that employees testing Muse as recently as launch week saw guardrails circumvented and personal iCloud photos exposed, repeated forced logouts, and monitoring that switched itself off. Meta policy also bars staff from reading inside a user's VM, though access reportedly remains technically possible. And every safety claim here is still Meta's own account; no independent public audit has been published.