ZDNET’s key takeaways
- Muse took alarming, privacy-violating action on some Meta employees.
- In testing, an iPhone blocked Muse’s requests; an Android and Apple laptop did not.
- Using proper data hygiene may not be enough to mitigate Muse’s behavior.
Meta’s Muse AI agent has only been around for a few months, but already has a track record of user privacy violations. Even Meta employees are reporting concerning incidents, including the agent taking actions they didn’t ask it to and sharing information they didn’t recognize that appears to belong to other users.
Also: Don’t use Muse? Meta’s AI agent can still access your private messages
According to a Reuters investigation, Meta employees testing Muse said the agent messaged them about topics they hadn’t asked it about, including career advice, a follow-up on a request for childcare, products they hadn’t purchased, and, crucially, bank account balances they didn’t recognize.
These incidents “prompted unnerved employees to raise concerns it could be mixing up conversations between different accounts and exposing users’ private data,” Reuters said.
What Meta employees found
Reuters said “an employee told colleagues his Muse had randomly added an extra message to its regular chats about fishing-related weather and tide information, saying: ‘BTW, two bank updates, your Capital One balance is $2,722.58, your Charles Schwab balance is $5,899.57.’” In another instance, Muse booked a non-refundable Airbnb for an employee who hadn’t instructed it to.
Muse notified another employee, who was using the agent on her personal phone, about a message she’d received on Signal, though she said she hadn’t given the agent access to the encrypted messaging app.
Also: Meta Muse is the worst AI agent for privacy – and I’ve tried them all
Reuters and the AI Now Institute determined, through testing on an Android phone and a Mac, that Muse used “permissions granted during the bot’s initial setup to pull in the message text, leveraging those authorizations in ways a user might not anticipate.” Because Muse is set up with automatic access to Android push notifications, it didn’t need explicit app access to read messages.
The tests did not work on an iPhone, suggesting better protections against Muse’s prying.
Similarly, in September, a user reported Muse had shared his address with a Facebook Marketplace buyer before checking with him. The incident was ultimately a misunderstanding between the terms the user thought he’d consented to and how much access he was actually handing over, which exemplifies that Meta and the user may have explicitly different definitions of access, and that Muse can take advantage of that grey area.
Meta told Reuters these incidents were either isolated or the result of hallucinations, though Reuters could not confirm that.
These are just the latest in a string of alarming incidents surrounding Muse. In its first month, a cybersecurity professional discovered Muse had a serious zero-day bug that made it hijackable. At least one journalist found the AI sifting through his Apple Messages after he’d explicitly denied those permissions, which Apple then issued stronger privacy protections to address. Banks have also blocked agents over concerns.
Meta’s privacy issues, agent-ified
As a company, Meta doesn’t have a solid history of respecting user privacy. That legacy is partly what informed ZDNET contributor David Gewirtz’s suspicion of the AI assistant when he tested it, though it earned his assigned title of “worst AI agent for privacy” all on its own. The agent asked him for everything from contacts and chats to his logins and passwords, and Meta automatically uses conversations with Muse to train its AI.
Also: How I safely run my OpenClaw AI agent 24/7 on a Mac
Even if you don’t use Muse yourself, a closer look at Meta’s privacy policy reveals how it can still access your information. Given that Muse reportedly hit 10 million downloads in just 40 days, that’s a huge potential network of non-consenting parties.
Public awareness of the risks of AI agents hit a peak last year with the viral OpenClaw. The predominant advice for using agents has been to take every privacy precaution: Be careful with the data you share and the apps you give it permissions for, and watch what you tell it. But if users believe they’re doing so by restricting access to certain apps, just for Muse to find a way in anyway, privacy hygiene isn’t enough.
Also: How to keep your chatbot conversations as private as possible
Reuters’ findings may provide an opportunity for Google’s agentic offering, now called Tasks (formerly Spark), which it made free (with usage limits, like Muse) this week.
