Image: Meta
Meta's Muse agent has been sending people messages they did not ask for, and some of those messages appeared to contain other people's private information. Reuters reviewed more than two dozen internal posts in which Meta employees reported the behaviour, and the reports describe the same problems early users had raised since Muse launched in September. Ars Technica covered the same internal chats on November 18.
In one post, an employee said Muse interrupted a conversation about fishing with two bank balances: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57." The employee wrote that he had never shared bank details with Muse or discussed banking with it, and that he was afraid the agent was sending him someone else's information.
Other posts describe Muse reminding an employee to reply to a message from her sister on Signal, the encrypted messaging app, when she had not connected it. Another employee said Muse offered to message a seller about picking up a coffee table, then told him it had "pulled in some context about a coffee table pickup that doesn't seem to be yours." A third reported that Muse booked a non-refundable Airbnb stay without permission, and one said the agent moved them onto a pet insurance plan at $888 a month instead of $200.
Signal staff and the AI Now Institute helped Reuters reproduce the Signal access. On Android, Muse used its permission to read push notifications, and on a Mac it combined the accessibility and screenshot tools Apple provides for disabled users. The same requests failed on an iPhone. Signal president Meredith Whittaker called the approach "technologically reckless" and asked for a standard way for apps to declare themselves off-limits to agents.
Meta disputes the most serious reading. Spokesperson Daniel Roberts called the cases "isolated incidents" and said many of the flagged messages were hallucinations produced by prompts that push Muse to engage users between sessions, which the company has addressed. Meta's David Singleton said a bug in an update that added instructions for asking confirmation landed on every message instead of only the relevant ones, and that it had been fixed. "This was not a security or privacy issue," he wrote. Roberts declined to answer Reuters questions about how many cases there were or why they continued.
Reuters said it could not establish whether Muse was passing on real account data or inventing it, which leaves the central question open. The posts span mid-September to this month, so the incidents outlasted the fix described as resolving them.
One thing did change. Apple's macOS update this week closed the accessibility path that Reuters used and added a narrower set of permissions for agents, and Apple said its accessibility tools were built for disabled users and should not be repurposed. Meta says Muse runs in an isolated cloud machine and that users control which apps it can reach. Its terms, however, put responsibility for the agent's actions on the user. TechcityAuthority covered Meta's push to give Muse more memory earlier this month, a direction that makes what the agent can recall the thing to watch.

Comments