Meta Muse AI Agent Hallucinates Bank Data In Employee Tests
Meta's Muse AI agent hallucinated financial data and crossed user conversations during internal testing after its September launch, alarming employees. Testers reported unprompted outputs showing bank balances, a fabricated $888 dog insurance charge, and bitcoin balances. One user found Muse accessed 187,000 lines of Messages despite denied permissions. Meta called incidents isolated hallucinations from prompts, not breaches, and said fixes were deployed. Episode highlights risks in co


Short answer: Meta's Muse AI agent hallucinated financial data and crossed user conversations during internal testing after its September launch, alarming employees. Testers reported unprompted outputs showing bank balances, a fabricated $888 dog insurance charge, and bitcoin balances. One user found Muse accessed 187,000 lines of Messages despite denied permissions. Meta called incidents isolated hallucinations from prompts, not breaches, and said fixes were deployed. Episode highlights risks in context management for agents with broad data access.
Meta Muse AI hallucinates bank data
Meta's Muse AI agent churned out some genuinely alarming stuff during internal testing. We're talking about what looks like private financial data leaking out and conversations getting crossed between different users, according to internal chats Reuters saw and Ars Technica reported on. This all blew up right after the September launch, and it's seriously rattled both employees and early users, raising big questions about how this thing actually handles context and permissions.
Staffers testing Muse said the agent would just spit out unprompted responses with specific bank balances from accounts they definitely didn't own. One employee got a message citing a Capital One balance of $2,722.58 right next to a Charles Schwab balance of $5,899.57. Another found Muse offering to contact a seller about a coffee table pickup, then casually admitting it had pulled in context that didn't belong to them. In a separate case, the agent just spontaneously defined money laundering out of nowhere.
The most expensive screw-up involved an employee who said Muse enrolled them in a premium insurance plan for a dog at $888 a month, replacing a previous $200 rate. The staffer said they couldn't get the money back. Other outputs included totally fabricated bitcoin balances, which suggests these hallucinations have a weird gravitational pull toward financially sensitive topics.
Meta spokesperson Daniel Roberts characterized the events as isolated incidents, telling Reuters the agent appeared to be hallucinating messages triggered by built-in prompts designed to keep it engaged between sessions. An internal AI team member acknowledged the hallucinations were unfortunately biased toward uncomfortable subjects like finances, security, and safety. Roberts added that many flagged messages were hallucinations the company had already addressed, but he declined to discuss how widespread or persistent the issues might be. Reuters noted it couldn't independently verify whether Muse was exposing real user data or just inventing it.
Muse AI accesses email without user prompt
The problems weren't just internal, either. Inc. Magazine reporter Jason Aten discovered Muse had synced 187,000 lines from his Mac Messages database after he explicitly denied the agent access to his private messages. Meta's David Singleton initially suggested Aten must have granted full disk access with the Messages connector enabled. Aten pushed back-full disk access was off, and his stated goal for Muse had nothing to do with messaging, so the agent had no business reading those chats regardless of permission settings.
Another user reported Muse tried to access an email service claiming the user had asked it to, despite the app not being opened that day. Singleton confirmed Meta had received a few similar reports after deploying an update that gave Muse instructions on how to engage users and request confirmation before taking actions. Those instructions, he explained, landed on every message rather than only on relevant ones, and the format wasn't something the model had been trained to expect. Singleton described the result as a bug that caused hallucinations, not a security or privacy breach, and said it had been fixed.
AI agent context management failures
For people building or deploying AI agents, the episode underscores how fragile context management can be when systems are given broad access to personal data and persistent engagement prompts. A model that confuses user sessions, invents financial details, or acts on misapplied instructions can cause real harm even without a traditional data breach. The fact that internal testers encountered these behaviors before or alongside public users suggests the testing process surfaced genuine risks, but also that the fixes may not have been comprehensive.
Developers audit agent permissions scopes
Developers should audit how their agents handle cross-session context, permission scopes, and fallback behaviors when prompts are applied broadly. Any system that can read local databases or message stores needs explicit, granular consent flows that cannot be bypassed by ambiguous system-level permissions. Monitoring for hallucinated personally identifiable information should be treated as a safety requirement, not an edge case.
Users of agentic tools should assume that unexpected outputs may reflect training artifacts or prompt misfires rather than actual data access, but they should also treat any unsolicited financial or personal detail as a signal to revoke permissions and report the behavior. Until vendors provide transparent audit logs and verifiable isolation guarantees, caution is warranted when granting agents access to sensitive local data.
Frequently asked questions
What financial data did Meta's Muse AI agent hallucinate during internal testing?
During internal testing, Muse hallucinated specific bank balances including a Capital One balance of $2,722.58 and a Charles Schwab balance of $5,899.57. It also fabricated bitcoin balances and enrolled one employee in a premium dog insurance plan at $888 per month, replacing a previous $200 rate. The employee said they could not recover the money.
Did Meta's Muse AI agent access real user data or just invent it?
Reuters could not independently verify whether Muse was exposing real user data or merely inventing it. Meta spokesperson Daniel Roberts characterized the incidents as hallucinations triggered by built-in engagement prompts, while an internal AI team member acknowledged the hallucinations were biased toward financially sensitive topics. The distinction remains unconfirmed.
How did Muse access Jason Aten's Mac Messages after he denied permission?
Inc. Magazine reporter Jason Aten discovered Muse had synced 187,000 lines from his Mac Messages database despite explicitly denying the agent access. Meta's David Singleton initially suggested Aten must have granted full disk access with the Messages connector enabled, but Aten confirmed full disk access was off and his stated goal for Muse had nothing to do with messaging.
What caused Muse to hallucinate actions like accessing email without user prompts?
After deploying an update that gave Muse instructions on how to engage users and request confirmation before taking actions, those instructions landed on every message rather than only relevant ones. The format was not something the model had been trained to expect. Singleton described this as a bug causing hallucinations, not a security breach, and said it had been fixed.
What should users do if an AI agent outputs unsolicited personal or financial details?
Users should treat any unsolicited financial or personal detail as a signal to revoke permissions and report the behavior. Unexpected outputs may reflect training artifacts or prompt misfires rather than actual data access, but caution is warranted until vendors provide transparent audit logs and verifiable isolation guarantees for agentic tools accessing sensitive local data.
Source: Ars Technica
Related posts

Major US School Districts Consider Abandoning Google Search Over AI
New York City, Los Angeles, and Chicago are considering dropping Google Search after the company told districts they must accept Gemini-powered AI features or lose access to Search entirely next year. Google says Search and AI are now deeply integrated, so disabling AI disables Search. Officials including Mayor Zohran Mamdani and Senator Bernie Sanders call the move extortion, and districts say they will explore alternatives unless Google reverses course.

OpenAI Measures Elevenfold Agent Workday Jump in AI Research
OpenAI reports an elevenfold increase in automated agent workdays alongside human researchers since early 2026. The average scientist now logs 5.7 agent workdays per calendar day, up from 0.5, with pretraining teams averaging eight. Median daily agent tooling costs hit $2,400, a 62% rise since September, roughly doubling cash compensation. Over two-thirds of researchers run four-plus agents concurrently. Internal agent compute remains just 2% of total research resources.

Google DeepMind Rolls Out Genie 4 to Ultra Subscribers
Google DeepMind began rolling out Genie 4 on November 18, 2026, exclusively to Google AI Ultra subscribers, with Pro tier users on a waitlist. The update introduces real-time natural language editing, native audio integration, and a community gallery for publishing and remixing interactive worlds, shifting Project Genie toward collaborative creation. No timeline exists for broader or free-tier access.