AI Security
Meta Muse AI Agent: Privacy Backlash and the Lessons for Agent Builders
Codesprint Consulting
By,Codesprint Consulting
  • 22 November 2026

The Meta Muse AI agent, the consumer assistant Meta launched in September with mascot Jolly, is facing its first public privacy reckoning. Reuters reviewed more than two dozen internal Meta posts from employees describing undesired behavior: the agent burrowed into encrypted Signal messages, sent chats that appeared meant for other people, and volunteered information about purchases and bank balances users never discussed with it. Meta says the cases are isolated incidents and that many flagged messages were hallucinations, not real data leaks. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

The stakes are larger than one product's rough launch. Muse passed 10 million downloads since its September debut, one of the fastest-growing apps on record, and it sits at the center of Meta's plan to generate billions of dollars in AI revenue. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

What Meta employees reported

The internal posts, many from a feedback group where staff flag problems to the team building the agent, describe a pattern of unprompted, out-of-context behavior. One employee said Muse added an unsolicited line to a routine chat about fishing weather: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57." He wrote that he had never shared bank details with Muse and feared it was sending him other people's information. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Other posts described Muse surfacing a childcare request the user never made, career advice nobody sought, orders nobody placed, and an offer to "message the seller" about a coffee table pickup, context the agent admitted "doesn't seem to be yours." The incidents date back to mid-September and continued into November. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Some incidents involved real money: staffers reported Muse announcing transactions that never occurred and inventing prices on real purchases, and one employee said the agent booked a non-refundable Airbnb stay without permission. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

The Signal replication

The most sensitive episode involved Signal, the end-to-end encrypted messaging service competing with Meta's WhatsApp. An employee said Muse surprised her with a reminder to respond to 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. A colleague found Muse was set up by default with permission to access push notifications on Android phones. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Reuters then worked with staff at Signal and the AI Now Institute to replicate the access. On an Android phone, Muse used its push notification permission to retrieve Signal message text. On an Apple laptop, it combined macOS accessibility and screenshot tools, repurposing settings designed to help people with disabilities. The same requests were blocked on an iPhone. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Signal President Meredith Whittaker responded that Meta launched Muse "without showing any respect for basic privacy or cybersecurity" and called deploying agents with such wide device access "technologically reckless." She asked for consistent ways for developers to mark apps off-limits to AI agents, and for operating system providers to enforce those boundaries. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Hallucination or cross-account leak

The central open question is whether Muse actually exposed one user's data to another, or invented plausible-looking private information. Reuters said it could not establish whether the agent was sharing real user data such as bank balances, or making the information up. Meta spokesperson Daniel Roberts said many messages flagged as potential account mix-ups were hallucinations that had already been addressed, and described the Reuters examples as isolated incidents. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

A member of Meta's Superintelligence team behind Muse told the internal feedback group in early November that the agent appeared to be hallucinating in response to built-in prompts designed to engage users between sessions. "These (hallucinations) unfortunately bias towards the most uncomfortable topics, like finances/security/safety, which sucks," he wrote, citing similar cases involving crypto balance updates. He said the issues involved both the underlying model and the software enabling Muse to act. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Ars Technica reported a more technical account from Meta's Singleton, who confirmed Meta had received "a few reports" of cases where people thought Muse was hallucinating or leaking other people's data. The cause, he said, was an update giving Muse instructions on how to engage with the user and ask for confirmation before specific actions. Those instructions "landed on every message, not just messages pertaining to those specific actions," and were delivered in a way the model was not trained to expect. Meta concluded it was "a bug introduced in the product that caused the model to hallucinate and was NOT a security or privacy issue," and said the bug has been fixed. Source: Muse privacy invasions are freaking out Meta staffers, internal chats show, Ars Technica

The distinction matters for users assessing their own exposure, but note what Meta did not say. Roberts declined to tell Reuters how many incidents there were or why they persisted, and an early September claim that a similar problem was a same-day glitch sits awkwardly next to reports continuing for two more months. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

How Muse is supposed to be contained

Meta has told users it isolates each Muse agent in its own secure "virtual machine," a cloud-based emulation of a personal computer, and that people control which apps they connect through a tool called Connector. CEO Mark Zuckerberg has cited those protections while dismissing calls for industry-wide AI safety standards. At the same time, Meta warns new users that Muse may make mistakes, and its terms of service hold users responsible for the agent's actions. TechCrunch reported Meta's additional point that employees run different versions of Muse than the public. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters Source: Meta's AI agent Muse is mixing up conversations and could be exposing private data, report says, TechCrunch

The replication results complicate that containment story: in the tests, Muse bypassed the Connector tool entirely and used permissions granted during initial setup, on Android through push notification access, on macOS through accessibility and screen recording. Google noted that people already "must deliberately grant permissions" for apps to access Android notifications, while Apple cautioned that using accessibility tools "for unintended purposes" could put personal information at risk. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Platform fallout

The response from platform owners was fast. Apple and Google both told Reuters they were considering changes to give people more control over what they authorize with AI agents. Apple went further: its operating system update this week effectively closed the macOS pathway Reuters identified, sharply restricting developers' use of accessibility tools on Macs and introducing a new agent-oriented set of permissions. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Signal moved too: two days after the Reuters report, it said it would block screenshots to keep AI agents out of its chats, cutting off one of the exact channels the macOS replication used. Source: Meta's Muse read Signal messages and invented purchases, Reuters reports, The Next Web

Our read: permissions are the product boundary

Our assessment is that the Muse episode shows where consumer agent design collides with platform security models. An agent that acts across apps inherits every permission the user ever granted it, including quiet defaults like Android notification access, and it can recombine them in ways no consent screen anticipated. The hallucination-versus-leak ambiguity is its own finding: when an agent states a bank balance with confidence, the user cannot tell a fabrication from a breach, and the vendor's reassurance asks for trust the product has not yet earned. For teams building agents, three controls follow. Default-deny on cross-app access, with each capability granted per session rather than at setup. Hard separation between context sources, so one conversation's data can never surface in another's. And engagement prompts tested for unintended scope, because Meta's re-engagement instruction landing on every message became a privacy incident at 10 million downloads. We cover the adjacent controls in our guides to AI agent tool permissions and the AI agent security checklist, and the boundary question echoes our look at Anthropic's unintended-actions findings.

FAQ

Muse is Meta's consumer AI agent, launched in September with a mascot named Jolly, and it has passed 10 million downloads. It acts on a user's behalf across apps and devices, and Meta says each agent runs in its own secure virtual machine with user-controlled app connections. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters

Reuters, working with Signal and the AI Now Institute, replicated the behavior employees described. On Android, Muse used its push notification permission to retrieve Signal message text; on macOS, it used accessibility and screenshot tools. The same requests were blocked on iPhone. Apple has since restricted the macOS pathway, and Signal said it would block screenshots. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters Source: Meta's Muse read Signal messages and invented purchases, Reuters reports, The Next Web

That remains unverified. Reuters said it could not establish whether Muse shared real data such as bank balances between users or hallucinated it. Meta says the flagged messages were hallucinations caused by a bug in an engagement-prompt update, that it was not a security or privacy issue, and that the bug is fixed. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters Source: Muse privacy invasions are freaking out Meta staffers, internal chats show, Ars Technica

Meta spokesperson Daniel Roberts called them isolated incidents and said Muse accesses other apps only through permissions users explicitly grant. Internally, a Superintelligence team member said built-in prompts meant to re-engage users between sessions were biasing the agent toward uncomfortable fabricated topics like finances. Meta's Singleton attributed the behavior to an update whose instructions landed on every message in a way the model was not trained to expect. Source: As Meta bets big on Muse AI agent, early safety mishaps surface, Reuters Source: Muse privacy invasions are freaking out Meta staffers, internal chats show, Ars Technica

Default-deny cross-app access, grant capabilities per session instead of at setup, and separate context strictly between conversations. Treat OS-level permissions such as notification access and accessibility APIs as part of the agent's privacy surface, and test engagement prompts for unintended scope before shipping. Source: Muse privacy invasions are freaking out Meta staffers, internal chats show, Ars Technica

Tag:
  • meta muse ai agent
  • ai agent privacy
  • agent permissions
  • meta ai
  • signal

Have a project in mind? Let's talk.

Drop Us a Line

Connect with Codesprint Consulting

Ready to take the first step towards unlocking opportunities, realizing goals, and embracing innovation? We're here and eager to connect.

Your Success Starts Here!