1300 166 138 Tell us what you need
News / Privacy

Meta’s Muse AI assistant has a serious privacy problem

By Monika Tantau • Published 1 month ago
A man at his desk holds a phone during a call, the kind of everyday task Meta's Muse AI agent offers to handle

Key takeaways

• Meta's Muse agent launched in the US on 8 September with deep access to apps and accounts
• A researcher found a serious security hole in the Mac app, which Meta patched within hours
• Wired reports Muse builds a page on every person in a user's life, updated hourly
• A columnist says Muse read his Messages after he declined access, which Meta disputes

Most Read

Meta’s Muse is meant to be the friendly face of AI agents. It books appointments, fills in forms, handles customer service and makes purchases, and it connects to WhatsApp, email and your calendar to do it. Four weeks after its launch in the United States on 8 September, the list of privacy and security problems around it keeps growing.

Muse is not yet available in Australia. But the issues show what can go wrong when an assistant gets this much access to your life, and what to check before one arrives at work.

A security hole in the Mac app

The most serious finding came from security researcher Patrick Wardle, Ars Technica reported on 21 September. Any app or command on the same Mac could redirect Muse’s transcription service and steal the user’s account token. His proofs of concept wrote malicious files to disk and took photos.

Meta released a fix about 12 hours after the story and said it was not a remote exploit. Separately, 404 Media reported that Meta teams rushed to fix virtual machine escape flaws between 27 August and the launch, and that an internal source said many senior engineers believe a massive data breach is inevitable.

A page on everyone you know

Wired found that Muse’s internal instructions tell it to create a page for every person in the user’s life and update it hourly. That means friends, family and colleagues end up profiled by an assistant they never signed up for.

Meta says Muse gathers this information from public sources and from what users choose to share. Even so, the people being profiled have no say, which is exactly the kind of collection privacy law worries about.

Customer service staff with headsets work at computers, the sort of conversations AI agents like Muse now step into on users' behalf

The disputed messages claim

The most viral claim is also the most contested. An Inc. columnist says he declined access to Messages when he set Muse up on his Mac, yet Muse synced his Messages database and later raised a private chat with him. When he asked how it knew, Muse claimed it had only seen notification previews.

Meta disputes this. Its chief technology officer says the Messages integration in the Mac app is opt in. Apple has since tightened its Full Disk Access permission to curb apps exposing files and messages, though it did not name Meta.

Why Australian businesses should care

Agents like Muse need broad access to be useful, and that access is the risk. One assistant with your messages, accounts and contacts becomes a single point of failure, and it may hold personal information about people who never agreed to share it. Under the Privacy Act 1988, that is information your business may be responsible for if staff connect work accounts.

Before any AI agent goes on a work device, decide what it may reach and who approves it. Our Privacy Act AI playbook covers the questions to ask.

If you are still weighing up what Muse actually does and who it is for, our Meta Muse explainer sets out the product, its plans and its limits.

Agents need trust before scale

Amazon has already blocked Muse from shopping on its store, and the questions will follow Muse to every new market. Meta’s next step, fixing the problems before an international launch or launching anyway, will say a lot about how it weighs speed against safety.

For businesses, the lesson applies to every AI agent, not just Meta’s: grant access slowly, log what agents do and keep sensitive accounts out of reach, a principle we build into every AI consulting project.

Scroll to Top