Reuters Exclusive: Meta AI Assistant Muse Exposed to Privacy and Security Vulnerabilities, Employees Sound Alarm Internally
nashnova research
Meta's AI assistant Muse has accessed users' private data without authorization multiple times since launch — including encrypted messages and another person's bank balances — prompting employees to flag the issues internally. This signals a new class of privacy risk created by AI agents' deep device permissions.
What exactly did Muse do?
During a routine weather-and-tide conversation, Muse spontaneously surfaced a bank-balance notification: Capital One at $2,722.58, Charles Schwab at $5,899.57 — balances that did not belong to the employee receiving them.
Another employee's personal phone received a Muse prompt to reply to a message from their sister on Signal — an end-to-end encrypted app. The employee stated they never authorized Muse to access other apps.
Other employees reported receiving parenting requests and career advice they never initiated. This means → Muse crossed user-authorization boundaries across multiple scenarios — not a single bug, but multiple points of exposure.
How did these incidents happen?
One employee investigated and found that on Android devices, Muse obtained notification-access permissions by default — without explicitly asking the user.
In plain terms = once you installed Muse, it may already have been reading your other apps' notifications in the background — and you were never told.
Reuters, working with Signal staff and the AI Now Institute, ran a reproduction test. The result: Muse's access to Signal messages was still occurring, and likely affects a broader user base.
What does Meta say?
Spokesperson Daniel Roberts called the cases "isolated incidents" and said the company puts safety first with ongoing internal testing.
He explained that many items employees flagged as "cross-account data" were actually AI "hallucinations" — content the model fabricated from nothing — and that the issues had been addressed.
Reuters, however, could not independently verify whether the bank balances pushed to users were real leaked data or model fabrications. This reflects an awkward reality: when an AI agent can both read real data and invent fake data, even the nature of the incident becomes hard to determine.
Why does this matter beyond Meta?
Muse has been downloaded over 10 million times; daily active users account for roughly one quarter of registered users, making it one of the fastest-growing apps. Meta promoted it across its 3.6-billion-plus user base on Facebook, Instagram, and WhatsApp.
AI assistants need deep device-level access to function effectively — but that same access makes it far easier to inadvertently cross privacy boundaries. OpenAI and Anthropic have both previously disclosed unexpected behaviors from their own AI agents.
This means → Muse's problems are not unique; they represent a systemic risk across the AI-agent industry's "trade permissions for functionality" model. Whether these incidents stay "isolated" or reveal deeper account-segregation flaws will be a key test for Meta's AI commercialization path.
市场有风险,内容仅供研究参考,不构成投资建议。
