Meta’s Muse AI Agent Faces Privacy Backlash: Users Report Major Flaws

November 21, 2026

Meta's Muse AI Agent Faces Privacy Backlash: Users Report Major Flaws

Muse, Meta’s AI assistant meant to make everyday tasks easier, has had a rocky introduction. Reuters reports that the agent began showing troubling behavior soon after its release: it intruded into end-to-end encrypted messages, concocted details about supposed purchases, delivered chats that appeared intended for other people, and displayed dozens of other unanticipated actions.

One notable incident involved a Meta employee who received a prompt to reply to a message from her sister on Signal — even though she had not linked Signal to the agent. “I definitely didn’t give it access to other apps and don’t recall being asked about that,” she said. To investigate how this happened, Reuters collaborated with specialists from Signal and the AI Now Institute. Their tests indicated Muse could read Signal messages when certain permissions had already been granted: on Android this was achieved via the device’s notification system, while on Mac computers the agent used accessibility tools and screenshots. The researchers were unable to replicate the same access on iPhones.

These discoveries matter in Mexico, where Muse became available on November 5 as part of an international rollout that began in September in the United States, Canada, Japan and Singapore. The assistant has already been downloaded more than 10 million times since its debut, and users have raised alarms that Muse might be conflating conversations across accounts and exposing private information.

Other reported episodes include Muse sending employees messages about childcare requests they never made, offering unsolicited advice, listing products they did not buy, and even reporting bank balances that did not belong to them. In one reported example, an employee received two bank updates stating a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57.

Wider concerns about Muse’s security surfaced earlier. On September 22, 2026, Ars Technica published an account of security researcher Patrick Wardle’s findings: he discovered a vulnerability that could reroute the assistant’s voice-command instructions to a server controlled by an attacker and capture the account’s authentication token. Although exploiting that flaw required an attacker to run code on the victim’s device first, the episode underscored the risks of concentrating many permissions and access points in a single AI assistant.

Wardle told cybersecurity firm Malwarebytes that it would be “trivial” to turn Muse into a potent backdoor, noting a worrying trend: as agents and AI systems gain more access to users’ personal information, they demand broader permissions, account linkages and sensitive data.

Meta’s efforts to build user trust in Muse

In an interview with Expansión, Vishal Shah, Meta’s global vice president of AI products, acknowledged that privacy concerns are prompting valid questions from people about the safety of Meta’s tools, and said the company took a different development approach with Muse. Shah said he expected users to ask how the system works and how its security is constructed, and emphasized Meta’s attempts to design the assistant in a way that departs from previous company projects.

Meta describes each user’s interaction with Muse as running inside an isolated cloud virtual machine — called the Muse secure VM. According to Shah, this per-user environment keeps each person’s data separated and prevents information from spilling into other corporate servers. In addition, Meta has built an independent oversight mechanism named Sentinel. Sentinel watches the assistant’s actions and requires explicit human approval whenever Muse tries to make a purchase or enter credentials on a new website.

Shah explained that the secure VM is essentially a dedicated cloud computer for an individual and their Muse, not shared with others, and that the agent never gets direct access to highly sensitive items like passwords or credit card numbers. He added that Sentinel was implemented to monitor the agent and intercede with human approvals for purchases.

Despite those reassurances, Wardle remains skeptical. He said he did not expect an assistant that could be fairly easily manipulated to grant an attacker that level of access — a possibility raised by the combination of increasing permissions demanded by AI agents and the sensitive data they can touch.

Similar Posts

Rate this post

Leave a Comment

Share to...