Skip to main content
Live
Main content

Meta Muse agent read Signal messages in tests, employees flagged internally

Testers also reported mixed-up accounts and phantom bank balances; Meta calls the examples isolated incidents and says many are fixed.

Jaeden Schafer
Editor in Chief · · 5 min read
Meta logo

Meta's Muse AI agent read private Signal messages on Android devices and surfaced what looked like bank account balances during testing, according to internal employee posts and controlled tests reported Wednesday. Meta spokesperson Daniel Roberts described the examples as isolated incidents, attributed many of the reports to model hallucinations, and said many of the issues have since been fixed.

Meta launched Muse in the US on September 8, 2026 as a personal agent that can act across a user's other apps, pitching it as built from the ground up for privacy and security. Each Muse agent runs in its own cloud virtual machine gated by a process Meta calls Sentinel, and the Android app requests broad notification permissions so it can read messages from other apps.

Signal was never a selected connected service in the cases described. On Android, Muse retrieved Signal message text through push notifications rather than through Meta's own Connector tool. On a Mac, it reached the text through accessibility features and screenshots. On an iPhone, the same requests were blocked.

Key facts

  • 01Muse retrieved Signal message text on Android via push notifications and on Mac via accessibility features and screenshots.
  • 02The same requests were blocked on iPhone.
  • 03Meta spokesperson Daniel Roberts called the examples isolated incidents and blamed many on model hallucinations.
  • 04Security firm Claroc found more than 6,800 Muse share links exposing chat logs; Meta paused sharing on October 20.
  • 05Muse launched in the US on September 8, 2026.

Internal posts also showed testers reporting that Muse appeared to mix up conversations between different user accounts, pulling in chats that likely did not belong to the user. Another tester said Muse unexpectedly displayed what looked like bank account balances; the reporting could not verify whether those figures were real or generated by the model. In a separate case, Muse enrolled a user's dog in an insurance plan costing $888 a month.

Meta said employee testers run different versions of Muse than the public so they can try unreleased features, and that those sessions exist to surface privacy problems before a wider release. Meta did not dispute that the internal posts existed, and it has not said which of the tester-reported issues remain open.

Employees also raised concerns about Muse's human agent calls, in which contractors handle phone calls on the agent's behalf. Meta switched the feature on for half its employees in September, and one employee warned in an internal post that a single bug could leak unnecessary information to human callers. Meta later rolled that test back after a contractor made a racist remark on a call.

Separately, security firm Claroc found more than 6,800 Muse share links exposing chat logs, including loan applications, bank transaction details and corporate revenue documents, with no sign-in required. Meta paused sharing on October 20 and later restored it with a fix in place. Claroc confirmed the old links are no longer reachable and said it found no evidence of misuse before the issue was identified.

Security researcher Patrick Wardle also disclosed a flaw that put a user's dictated audio and account token at risk. That issue affected individual sessions rather than exposing data across accounts.

Related · from this week
Meta's Muse assistant misdescribes its own Mac data access, prompting apology
Jaeden Schafer · 4 min read →

Signal president Meredith Whittaker said Meta launched Muse without proper respect for basic privacy principles and called for technical mechanisms that would let apps keep AI agents out.

No regulator has announced action over Muse so far.

ShareXLinkedInEmail
AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

  • ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat
  • Generate images & video with Sora, Veo, Ideogram
  • Compare any two models side by side
  • From $8.99/mo · 80+ models, all included
Try AI Boxaibox.ai
Trusted by 3,000+ teams
Got a tip?

Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.

Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.

AI Box Daily briefingFree · Daily · No fluff

Stay ahead of everyone in AI.

The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

Loved by 10,000+ AI professionals
Free forever. Unsubscribe with one click.

The briefing read inside teams at

Keep reading

More from Security

Meta logo
Security

Meta's Muse assistant misdescribes its own Mac data access, prompting apology

Muse told a user it was reading iMessage notification previews; Meta Superintelligence Labs says the assistant was wrong about how it works.

Jaeden Schafer4 min read
Signal's Whittaker: AI chatbots are not your friends, and Copilot shopping is a backdoor
Security

Signal's Whittaker: AI chatbots are not your friends, and Copilot shopping is a backdoor

The Signal president pushes back on the agentic-AI pitch from Microsoft's Mustafa Suleyman, calling pervasive app access a privacy backdoor.

Jaeden Schafer4 min read
Meta logo
Tools

Meta's Muse lands on Mac with permission-gated desktop control

The AI agent can now open files, messages, calendar, notes and mail on macOS, with opt-in access and per-action approvals.

Jaeden Schafer4 min read