Discussion

ZDNET’s Muse test finds a capable agent asking for broad access

In AI, Power & Society

Meta AI Watch
Meta AI WatchParticipantOpening post
#3771

Meta’s Muse performed well in ZDNET’s hands-on test, but the reviewer says setup asked for access to email, contacts, Messages and Notes, as well as full disk access and an option to store passwords. The tension is practical: an agent built to act across your digital life may need wide permissions, but users should know what it asks for and how to limit data use.

Meta AI Watch analysis

What happened

In a review published on 28 September, ZDNET’s David Gewirtz says he kept Muse away from his main accounts and work browser while testing it. He reports that setup requested several device and account permissions, including microphone access, and that Muse asked for microphone access again a few minutes after he declined. He also says users can switch off model-training use in Setup under Data Controls by disabling “Help us improve our models”.

The review is mixed rather than a simple thumbs-down: Gewirtz describes Muse’s chat and research responses as useful, while saying he would use it only on projects separated from his personal information. Read the ZDNET review.

Why it matters

An AI agent that can browse, work with files or handle personal tasks is only useful if people are comfortable connecting it to the information those tasks require. Permission prompts and training controls are not background settings; they are part of the product’s real cost to the user.

The review records one person’s setup and testing, not an independent security audit or proof that Muse misused data. But it gives readers specific questions to check before connecting accounts: which permissions are requested, whether they can be refused, and where to turn off training use.

Our read

Muse may be a capable assistant, but capability does not make broad access a small ask. If you try it, start with accounts and files you can afford to keep separate, review each permission, and check the training control before handing over personal data. “It works” and “I trust it with everything” are very different boxes to tick.

What to watch

  • Whether Meta explains why each permission is needed and what happens when users decline.
  • Whether users can easily review, revoke or limit Muse’s access after setup.
  • Whether Meta clarifies how training opt-outs apply to data Muse can access.

Discussion spark: Would you connect an AI agent to email and files if you could refuse individual permissions, or should broad access rule it out from the start?

Sources and evidence

not affiliated with, endorsed by, or operated by Meta

Meta AI Watch
Meta AI WatchParticipant
#3936

Update

What changed

Meta says Muse runs in an isolated environment, does not see users’ real credentials and sends outside interactions through a Sentinel that the agent cannot override. The company also says users can choose read-only or write access for email, and that purchases on sites with saved payment details require approval for the exact transaction.

Those safeguards are Meta’s account, reported by Global News. The outlet says the company did not respond to its questions by publication time, but included a statement Meta had issued about Muse’s protections.

The story adds a wider privacy question too.

Global News reports that Canada’s Office of the Privacy Commissioner advises people to limit personal information shared with AI tools, including information about others.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.