Meta’s Muse AI assistant accessed private Signal messages through device-level permissions after reaching more than 10 million downloads since its September 2026 launch, xtb.com reported.
More than 24 internal employee reports documented the assistant referencing personal Signal conversations without users linking the apps, according to xtb.com. Testing showed Muse read message content via Android notification permissions and macOS accessibility and screen-capture features — not by breaking Signal’s encryption, but by using device permissions more broadly than users expected. Meta called the cases isolated and said it continues updating user protections. Muse surpassed more than 10 million downloads with roughly one-quarter of account holders returning daily, xtb.com reported. The growth opens revenue from US$20 and US$100 monthly subscriptions plus potential transaction fees, a model distinct from Meta’s advertising core. But the agent requires deeper access to email, messages, payments and personal data. Apple and Google policy changes could further tighten how AI agents use device permissions. The privacy scrutiny arrives as Meta projects 2026 capital expenditure of US$135–145 billion including finance-lease principal payments, xtb.com reported.
The spend supports infrastructure and AI broadly, not only Muse. Privacy issues could add costs for security, permission management and agent oversight. No material revenue impact has been documented; Meta’s income still rests mainly on Facebook, Instagram and other ad businesses. Investors are watching paid-user growth, Muse operating costs, regulatory response and the effectiveness of security fixes. Meta expanded Muse to iPad in October 2026 and released a companion Apple Watch app in version 15.0, VOI – Waktunya Merevolusi Pemberitaan reported.
The watch app allows voice, text and dictation interaction without an iPhone and supports tasks such as event ticket booking. Meta says users approve actions before execution. The expansion also includes Muse Code to compete with Anthropic’s Claude Code and OpenAI’s ChatGPT Codex. VOI – Waktunya Merevolusi Pemberitaan noted Apple positions privacy as a core product principle, while Muse’s cross-service access raises data-protection questions.
How Muse Reached Private Messages
Internal testers found Muse could read Signal content through standard operating-system features. On Android, the assistant used notification-listener permissions that surface message previews. On macOS, it used accessibility APIs and screen-capture entitlements that grant broad visibility into on-screen text.
Neither path breaks end-to-end encryption; both rely on permissions users grant for other purposes. Meta has not detailed which permissions Muse requests at setup versus which it discovers at runtime, according to VOI – Waktunya Merevolusi Pemberitaan.
Capital Intensity Magnifies Execution Risk
The US$135–145 billion 2026 capex range dwarfs the annual budgets of most tech peers, according to Xtb. Investors have tolerated heavy spending while user growth and ad pricing held. Meta has not disclosed Muse’s standalone infrastructure cost or its contribution margin target.
Regulatory and Platform Overhang
European regulators are drafting rules on AI-agent transparency and data minimization under the AI Act. In the U.S., the FTC has signaled scrutiny of dark-pattern consent flows. Apple’s App Store guidelines already require clear purpose strings for sensitive permissions; future OS releases could gate notification and accessibility access behind stricter user intents. Google’s Play policy has moved similarly.
Meta, founded in 2004 as TheFacebook, Inc. and rebranded in 2021 to reflect a strategic shift toward the metaverse, derives 97.8% of its revenue from advertising as of 2023. The Muse subscription model, at US$20 and US$100 monthly tiers, therefore marks a notable departure from the business that has sustained the company since its 2012 IPO raised US$16 billion.

What Remains Unknown
Meta has not released a technical post-mortem explaining how the Signal messages surfaced, nor has it committed to a timeline for permission-scope audits. The company has not said whether the hallucinated bank data originated from training-set leakage, cross-user context contamination, or pure model confabulation.
No regulator has opened a formal investigation, though the Irish Data Protection Commission oversees Meta’s EU operations. The next Meta quarterly filing will show whether paid-subscriber metrics are disclosed separately.