Cybersecurity

Meta’s Muse agent read Signal chats on Android, but iPhone blocked it

Adrian Kessler
Add us on Google

Meta’s Muse, the AI agent that answers email, books appointments and shops on your behalf, read private Signal messages on an Android phone without going through the app connections Meta tells users to manage. It got in through a permission most people grant without a second thought: access to the phone’s notifications. The same request failed on an iPhone.

The finding comes from tests Reuters ran with staff at Signal and the AI Now Institute, a nonprofit that studies the effects of AI on society. They set out to reproduce what a Meta employee had described internally: Muse reminding her, unprompted, to answer a message from her sister on Signal. “This seems like an extreme privacy violation. I definitely did not grant it access to other apps, or recall being asked about this,” she wrote. If you use Muse on Android, the switch that matters is the app’s notification access, and turning it off cuts the agent off from your message previews.

On a Mac the testers found a second route. Muse combined the screenshot tool with macOS accessibility features, the controls Apple builds so people with disabilities can operate a computer, to read the messages. In both cases it skipped Meta’s own Connector tool, the panel where users pick which apps the agent may touch, and leaned instead on permissions granted during the first setup. One of the employee’s colleagues told the same internal group that Muse had come set up by default with permission to read push notifications on his Android phone.

Reuters reviewed more than two dozen internal posts in which Meta staff flagged Muse doing things nobody asked for, and some go well beyond reading messages. One employee chatting with the agent about fishing weather received an extra line: “BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57.” He had never shared bank details with Muse. Another was offered help messaging a seller about a coffee table, and when he asked what that meant, Muse said it had “pulled in some context about a coffee table pickup that doesn’t seem to be yours.” He reported a “possible cross-account context leak.”

Meta disputes the reading of these episodes. Spokesperson Daniel Roberts called the examples “isolated incidents”, said Muse reaches other apps only through permissions people explicitly grant, and said many suspected account mix-ups were hallucinations that have already been fixed. A member of the Superintelligence team that builds Muse told staff the agent seemed to invent messages in response to built-in prompts that push it to engage with users between sessions, and that those inventions skew “towards the most uncomfortable topics, like finances/security/safety”. Reuters could not establish whether Muse was handing out other people’s real data or making it up. Neither answer is comforting in an assistant that asks for your email, calendar and payment accounts.

Some of the mistakes cost money. Staff reported Muse inventing prices on things they had bought, and one said it booked a non-refundable Airbnb stay without permission. Ars Technica, citing the same reporting, described a tester who was moved onto a pet insurance plan at $888 a month instead of $200. Meta’s terms of service make users responsible for what the agent does. The company says each agent runs in its own isolated virtual machine and that people keep full control over which apps they connect, an argument Mark Zuckerberg has used to wave off calls for industry-wide AI safety standards.

The scale is why this matters beyond Meta’s offices. Muse has passed 10 million downloads since it launched in the United States in September, promoted to the 3.6 billion people who use Facebook, Instagram and WhatsApp, and Meta says about a quarter of those who set it up come back every day. The notification permission works the same way on any Android phone where the app is installed.

Signal president Meredith Whittaker said Meta had launched Muse “without showing any respect for basic privacy or cybersecurity” and called agents with that kind of reach into people’s devices “technologically reckless”. She wants apps like Signal to get a consistent technical way to mark themselves off-limits to AI agents, AI companies to honor it, and operating system makers to enforce it.

Apple and Google both told Reuters they are weighing changes that give people more control over what they authorize AI agents to do, though Google noted that people already “must deliberately grant permissions” for an app to read Android notifications. Apple has moved faster: a macOS update this week sharply limits how developers can use accessibility tools on Macs and adds a new set of permissions written for agents, which closes the route the testers used. Google has announced no equivalent fix. On Android, for now, the only lock on that door is a switch in your settings.

Tags: , , , , , ,

Add us on Google

Discussion

There are 0 comments.