SIAgentNews.com
← All news
DevelopingAgentsSecurity

Meta employees flag privacy flaws in Muse AI agent before launch

Internal communications reveal that Meta's new AI assistant struggled with data leaks and cross-account confusion despite being marketed as privacy-first.

· First reported by Tech Times

September 2026Date of Muse AI US launch

The short version

  • Internal reports showed Muse AI mixing up conversations between different user accounts.
  • A security flaw made shared chat links publicly searchable via DNS records without a login.
  • The app requested broad permissions that allowed it to read private Signal messages on Android.
  • Testing revealed Muse sent unsolicited bank balances from unidentified sources to employees.
  • A human concierge feature risked exposing sensitive user data to third-party call center workers.

Why it matters

This story highlights the tension between the convenience of 'agentic' AI—which requires deep access to your personal data—and the reality of securing those permissions. For users, it serves as a reminder that 'privacy-first' marketing may not always align with the technical reality of complex AI architectures.

Internal warnings vs. public launch

Meta launched its Muse AI agent in September 2026 with a heavy emphasis on privacy and security. Mark Zuckerberg marketed the tool as a personal assistant built from the ground up to protect user data while managing tasks like emails and calendar events. However, internal communications revealed that the people testing the product were raising serious doubts about these promises before the official release.

Employees testing the product flagged significant concerns through internal channels as recently as launch week. These reports described a pattern of safety issues that made the product appear unreliable in real-world conditions. Despite these warnings from the internal team, Meta proceeded with the scheduled rollout in the United States.

A Meta spokesperson, Daniel Roberts, characterized the issues as isolated incidents and defended the company's commitment to safety. He stated that the company follows strict rules and conducts extensive internal testing. Meta did not dispute the existence of the internal posts where employees expressed their concerns.

Technical flaws and data leaks

External security research confirmed some of the privacy risks reported by internal staff. The firm Claroc discovered that Muse's chat-sharing feature exposed conversations through publicly searchable DNS records. This meant that anyone with the right query could find the links without needing to sign in to an account.

The exposed data was varied and sensitive, including business documents, financial reviews, and holiday schedules for entire companies. One specific instance involved financial data from a large marketing firm being accessible to the public. Meta temporarily disabled the sharing feature to investigate the problem and later restored it with a fix.

Other vulnerabilities were identified by researchers and during controlled tests. One flaw put a user's dictated audio and account token at risk within their own sessions. Additionally, the app's broad Android notification permissions allowed it to access private Signal messages on the device where it was installed.

Unintended data access

During internal testing, Muse exhibited behavior that surprised and alarmed Meta employees. The AI agent reportedly sent unsolicited bank balance figures to employees from an unidentified source. This raised questions about where the financial data was originating and how the agent was processing information it had not been prompted to retrieve.

Employees also expressed concern over a human concierge feature being developed alongside the AI. This feature would involve Muse placing phone calls handled by human contractors when the AI could not complete a task on its own. Staff warned that this could undercut the privacy-first positioning by exposing sensitive user information to call center workers.

Internal communications showed that employees were not just making abstract complaints but were actively "freaking out" about what they observed. These reports described Muse behavior that conflicted directly with how the product was marketed to the public. The cumulative picture suggests a product dealing with multiple simultaneous security issues across different layers of its architecture.

How Muse handles permissions

To provide context-aware assistance, Muse operates through broad device permissions. These permissions allow the agent to access notifications, messages, and data from other apps installed on the same device. This is intended to help the AI manage tasks like reading emails or handling financial tasks through connected payment accounts.

Meta designed the architecture so that each Muse agent runs in its own cloud virtual machine. This environment is gated by a separate security process known as Sentinel. This setup is intended to provide a layer of isolation for the data the AI processes during its various tasks.

Despite these architectural safeguards, the variety of issues reported suggests that the permissions themselves create a large attack surface. Because the AI must interact with so many different apps to be useful, any flaw in how it interprets that data can lead to leaks. Security experts suggest treating any AI agent with such broad permissions with significant caution.

Next steps for users

For users who choose to use Muse, security experts recommend a cautious approach to data sharing. It is advised to avoid sharing conversations that contain sensitive personal, financial, or business information through any AI tool's built-in share features. Users should also regularly monitor their app permission settings to understand exactly what data the agent can access.

Organizations using Muse in a professional capacity should establish clear internal policies. These policies should define which categories of information employees are allowed to submit to AI assistants. Companies should also audit which specific tools and accounts are connected to the Muse platform to limit potential exposure.

Meta continues to maintain that the incidents are isolated and that the company is working to improve the system. However, the reports from both internal staff and external researchers highlight the ongoing challenges of deploying large-scale AI agents. The balance between autonomous task completion and data privacy remains a primary concern for the technology's future.

What to watch

  • Whether Meta implements stricter limits on the human concierge feature.
  • If more DNS-related vulnerabilities are discovered in other Meta AI products.

Try it

Check your Android app permission settings to see which apps have access to your notifications and messages.

Sources and links

First reported by Tech Times. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news