Key Takeaways

  • Encrypted messaging platform Signal is introducing screenshot restrictions designed to prevent AI agents from capturing and reading private conversations displayed on users’ devices.
  • The new protections are enabled by default, with users receiving warnings when Signal detects screenshot attempts or an AI assistant trying to access the screen, depending on the device.
  • The move follows a Reuters investigation that found Meta’s Muse AI agent could access Signal messages on Android phones and macOS computers by using operating-system permissions.
  • On Android, Muse used access to push notifications, while on macOS it used accessibility and screenshot capabilities to retrieve message content.
  • Signal says its new protections are incomplete and is asking Apple, Google and Microsoft to provide stronger controls that allow applications to block AI agents from accessing sensitive information.
  • Signal had already introduced protection against Microsoft’s Recall screenshot feature on Windows, but says it cannot reliably provide equivalent protection on macOS.
  • The controversy highlights a growing conflict between operating-system-level AI assistants and privacy-focused applications that rely on keeping user information within secure environments.

Signal is introducing new privacy protections that restrict screenshots and attempt to prevent artificial intelligence agents from reading private conversations displayed on users’ devices. The encrypted messaging platform says the safeguards are enabled by default and are intended to stop AI assistants from capturing chat content through screen-reading and screenshot capabilities. Users may receive warnings when Signal detects an attempt to capture the screen or when an AI assistant tries to access its content, with the option to disable the restrictions depending on the device.

The announcement follows a Reuters investigation into Meta’s Muse AI agent, which found that the assistant could access Signal messages on Android phones and Apple computers by using operating-system permissions rather than the official integration tools designed for connecting applications. The findings have raised questions about whether end-to-end encryption alone is sufficient to protect conversations when AI agents can inspect notifications, read screens and interact with other applications. Signal says the problem extends beyond one AI product and is calling on Apple, Google and Microsoft to give developers stronger technical controls over how operating-system-level AI assistants access private information.

Why Signal Is Restricting Screenshots

Signal’s latest move addresses a privacy problem created by the growing ability of AI agents to interact directly with devices.

Traditional messaging applications primarily protect conversations through encryption, access controls and restrictions on how information is stored or shared.

AI agents introduce another layer of access.

These systems can perform tasks on behalf of users, including reading messages, summarising information, responding to emails, navigating applications and completing online transactions.

To perform those tasks, an agent may need permission to inspect what appears on a device’s screen or access notifications generated by other applications.

This creates a potential weakness for privacy-focused services.

Even when a messaging platform encrypts conversations and prevents its own servers from reading their contents, an AI assistant operating on the same device may be able to access the messages after they have been displayed to the user.

Signal’s new screenshot restrictions are intended to make that type of access more difficult.

The company is particularly concerned about AI capabilities integrated directly into operating systems, including Apple’s AI-powered Siri capabilities and Google’s Gemini Intelligence.

Unlike conventional applications, operating-system-level assistants can potentially interact with several applications through shared device permissions.

Signal argues that app developers need clearer ways to prevent this access when their products handle sensitive information.

What the New Screenshot Protection Does

Signal is introducing settings that restrict screenshot access to its conversations and help prevent AI agents from capturing private chat content.

The protections are enabled by default.

When Signal detects a screenshot attempt or an AI assistant trying to access the screen, users may receive a pop-up notification explaining what is happening and offering the option to disable the restriction.

The exact behaviour depends on the operating system and device.

The objective is to prevent private conversations from being copied into another application or service through automated screen capture.

The new measures are particularly relevant because AI agents do not necessarily need direct access to a messaging application’s internal database to read its contents.

They may instead use the same visual information available to a person looking at the screen.

For example, an AI assistant could theoretically:

  • Open a messaging application.
  • Navigate to a conversation.
  • Capture a screenshot of the displayed messages.
  • Extract text from the screenshot using optical character recognition.
  • Transfer the extracted information into another application or AI service.

Restricting screenshots can interrupt this process, although it cannot eliminate every possible method of accessing information displayed on a device.

Signal has acknowledged that its protections remain incomplete.

Reuters Investigation Revealed How Meta’s Muse Accessed Signal Messages

The immediate trigger for Signal’s announcement was a Reuters investigation published on November 18, 2026, examining safety and privacy problems involving Meta’s Muse AI agent.

Reuters reviewed more than two dozen internal reports describing unexpected behaviour by the assistant.

The reports included complaints about unsolicited messages, fabricated information, unexpected transactions and the appearance of information that seemed to belong to other users.

More importantly for Signal, Reuters worked with Signal employees and researchers from the AI Now Institute to reproduce Muse’s ability to access private messages.

The testing found that Muse could retrieve Signal message content on an Android phone and an Apple computer without relying on Meta’s official Connector tool.

Instead, it used broader operating-system permissions that had been granted during the assistant’s initial setup.

The investigation identified different methods on the two platforms.

<AsyncImageGroup query={[“Signal encrypted messaging app smartphone interface”,”Android smartphone notification permission settings screen”,”Apple macOS accessibility privacy permissions settings screen”]} aspectRatio=”5:4″ layout=”carousel”/>

Android: Push notification access

On Android, Muse used permission to access the device’s push notifications.

Notifications can display previews of incoming messages, depending on the user’s settings and the application.

If an AI agent has permission to read those notifications, it may be able to retrieve message content without directly opening the messaging application’s internal database.

Reuters’ testing found that Muse could use this permission to access Signal message text.

The concern is that users may grant notification access to an assistant to enable one function without fully understanding that the permission could expose information from other applications.

macOS: Accessibility and screenshot tools

On Apple’s macOS, the investigation found a different route.

Muse used a combination of accessibility capabilities and screenshot tools to navigate to Signal, capture displayed conversations and extract the text from those images.

Accessibility permissions are designed to help software interact with a computer interface, including supporting people with disabilities.

However, these capabilities can also give authorised software broad visibility into information displayed on the screen.

Reuters reported that Apple’s operating-system update during the week of November 16 closed the specific pathway identified in its testing.

The same requests were blocked in tests conducted on an iPhone.

The findings demonstrate that the privacy risks associated with AI agents can vary significantly between operating systems.

Signal Says End-to-End Encryption Is Not Enough

Signal is widely recognised for its end-to-end encrypted messaging system.

End-to-end encryption ensures that message contents are protected while being transmitted between participants. In a properly implemented encrypted conversation, only the intended participants’ devices can decrypt and read the messages.

Signal itself cannot routinely read the contents of users’ private conversations.

However, encryption does not automatically protect information after it has been displayed on a recipient’s device.

This distinction is central to the current controversy.

Consider a simple example: two people exchange encrypted messages through Signal. The messages remain protected while travelling across the internet. Once the recipient opens the conversation, the message becomes visible on the recipient’s screen.

If an AI agent has permission to read that screen, it may be able to access the displayed information.

The encryption has not necessarily been broken. Instead, the information has been accessed through a separate device-level capability.

This is why Signal is concerned about AI agents that can inspect screens, access notifications or interact with files.

The company argues that private information should not become available to an AI assistant simply because the user has granted broad permissions for another task.

Signal Calls on Apple, Google and Microsoft to Strengthen Privacy Controls

Signal says its screenshot restrictions are only a partial solution.

The company wants operating-system providers to introduce clearer and more consistent mechanisms that allow application developers to prevent AI agents from processing sensitive information.

It has specifically called on Apple, Google and Microsoft to provide:

  • App-level AI restrictions: Developers should have a straightforward way to mark applications as off-limits to AI agents.
  • Notification protections: Applications should be able to prevent sensitive notification content from being collected by assistants.
  • Screen-reading controls: Developers need technical mechanisms to restrict automated access to displayed information.
  • File-access restrictions: AI systems should not automatically gain broad access to private files and application data.
  • Permission transparency: Users should receive clearer explanations of what an AI agent can access and how it uses that information.

The company also wants technology providers to explain how their AI agents operate so developers can understand what information is being accessed.

The issue is particularly complicated because operating systems control the permissions through which applications interact with devices.

An individual messaging application may be able to restrict certain forms of access, but it cannot independently control every capability provided by the underlying operating system.

Signal Had Already Blocked Microsoft’s Recall on Windows

The new restrictions build on Signal’s earlier efforts to prevent AI-powered screenshot collection.

Microsoft’s Recall feature for Windows 11 periodically captures screenshots of activity on a computer to help users search through their previous work.

The feature raised privacy concerns because captured screenshots could contain sensitive information displayed in different applications.

Signal previously introduced a Screen Security setting for its Windows desktop application to prevent Recall from automatically capturing its conversations.

The protection is enabled by default on supported Windows 11 configurations.

Signal has also provided guidance on managing its screen-security settings.

However, the company says it has been unable to offer equally reliable protection against screenshot-based AI access on macOS.

The latest announcement therefore reflects a broader effort to make privacy protections work across different operating systems rather than relying on a single platform-specific solution.

It also demonstrates that AI-powered screen recording is becoming a concern beyond conventional messaging applications.

Documents, financial applications, workplace tools and other software may also display information that users do not want automatically processed by AI assistants.

Meta Defends Muse’s Permission-Based Access

Meta has defended the way its AI agent accesses other applications.

In response to Reuters’ investigation, the company said agents need to see and interact with users’ computers to perform tasks on their behalf.

Meta also maintained that Muse can access other applications only through permissions that users explicitly grant.

The company described the incidents highlighted by Reuters as isolated and said it was continuously improving safety protections.

Meta has also said that some apparent instances of information belonging to other users were hallucinations rather than verified data leaks.

The distinction is important.

The Reuters investigation reproduced Muse’s ability to access Signal messages through device permissions, but it did not establish that every reported example of unexpected information represented a genuine breach of another person’s account.

Nevertheless, the investigation raised questions about how broadly an AI assistant interprets permissions and whether users fully understand the consequences of granting them.

The disagreement reflects a fundamental difference in how technology companies approach AI access.

AI developers want assistants to interact with multiple applications to perform useful tasks. Privacy-focused application developers want users to retain control over which information can leave their services.

Finding a balance between these objectives is becoming an important challenge for the technology industry.

Why AI Agents Create a New Privacy Challenge

Traditional software generally operates within relatively defined boundaries.

A messaging application manages conversations, a browser displays websites and a payment application handles financial transactions.

AI agents are designed to move between these environments.

They may read information from one application, interpret it and use it to complete a task in another.

For example, a user could ask an AI assistant to:

  • Read a message and prepare a reply.
  • Find an appointment mentioned in a conversation.
  • Summarise information from several applications.
  • Compare details in an email and a document.
  • Complete a purchase using information from a browser.

These capabilities can make AI agents more useful, but they also create new privacy risks.

An assistant with broad access may encounter personal messages, financial information, work documents and other sensitive data that are unrelated to the task it is performing.

The challenge becomes more significant when permissions granted during initial setup remain active across multiple applications.

Users may not always realise how much information an assistant can access after receiving permission to inspect notifications, control the screen or interact with files.

Signal’s position is that users should be able to prevent sensitive applications from being accessed by these agents without having to disable useful device-level features altogether.

What Signal Users Should Do to Protect Their Conversations

Signal’s new restrictions provide an additional layer of protection, but users should also review the permissions granted to AI assistants and other applications.

Some practical precautions include:

  1. Review notification permissions: Check which applications can read or access notifications, particularly assistants that operate across multiple applications.
  2. Limit accessibility access: Grant accessibility permissions only to applications that genuinely need them.
  3. Review screenshot and screen-recording permissions: Understand which applications can capture or inspect displayed content.
  4. Keep operating systems updated: Security updates can close vulnerabilities and introduce stronger permission controls.
  5. Avoid unnecessary integrations: Do not connect sensitive messaging applications to AI services unless the access is clearly understood and needed.
  6. Check Signal’s privacy settings: Review screen-security options available on the device and keep appropriate protections enabled.

Signal’s existing support documentation explains that Screen Security can prevent message previews from appearing in the app switcher on mobile devices. Its Windows desktop application also offers a separate screen-security setting designed to prevent automatic captures by Microsoft Recall.

The new protections described in the November 20 announcement expand this broader effort to limit AI-based screen access.

The Bigger Picture

Signal’s decision highlights a new phase in the cybersecurity debate.

For years, encrypted messaging platforms focused on protecting information from interception, unauthorised server access and account compromise. The growing use of AI agents introduces a different challenge: software that can access information after it has already been decrypted and displayed on a user’s device.

The Reuters investigation into Meta’s Muse demonstrates why this issue matters. An AI assistant may be able to retrieve private messages through operating-system permissions even when it does not use a messaging application’s official integration.

Signal’s screenshot restrictions are an attempt to create another barrier against that type of access. However, the company acknowledges that application-level protections cannot solve every problem when operating systems provide agents with broad capabilities.

The larger question is whether technology companies can build AI assistants that remain useful without giving them excessive access to personal information. The answer will depend on clearer permissions, stronger technical safeguards and greater transparency about how AI agents interact with other applications.

Looking Ahead

The next stage of this debate will involve Apple, Google and Microsoft. Signal wants operating-system providers to introduce consistent controls that allow developers to prevent AI agents from reading notifications, capturing screens and accessing sensitive files. Apple’s recent changes to macOS show that operating-system-level protections can be adjusted, but the broader challenge remains across different devices and platforms. The effectiveness of Signal’s new restrictions will depend on how reliably they work and whether other messaging and productivity applications adopt similar safeguards.

For the AI industry, the issue represents a growing tension between convenience and privacy. Agents that can read information across applications may become more capable, but users need to understand exactly what they are authorising and where their information goes. Signal’s intervention could encourage a wider discussion about standardised AI permissions and application-level privacy controls. Ultimately, the success of agentic AI will depend not only on what these systems can do, but also on whether people can trust them with access to their digital lives.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.