An employee using Meta’s Muse received a reminder to answer a message from her sister on Signal. She had not connected Signal through the agent’s usual integration system and said she did not remember giving Muse access to other apps. A reminder like that might be welcome from the messaging service itself, but coming from an assistant that was never supposed to read the conversation, it raised a rather different question: how did the agent know?
The account appeared in a Reuters investigation published on November 18, which examined more than two dozen internal reports about Muse. Reuters subsequently worked with Signal and the AI Now Institute to reproduce access to Signal message text. On Android, Muse could read the text through notification access granted during setup; on macOS, it could use accessibility and screenshot permissions. The equivalent requests were blocked in the reporters’ iPhone tests.
None of this required breaking Signal’s end-to-end encryption. Messages are decrypted for the person reading them, and parts of that content can appear in a notification or on a screen that another authorized component can inspect. The agent reached those representations of the information without using Meta’s Signal connector. That is the ordinary request behind a difficult technical problem: I want to make an application off-limits to an agent without having to discover every indirect route by which the agent might encounter it.
A personal agent needs access to do useful things. If I ask one to arrange a journey, it may need my calendar, booking confirmations, a browser and perhaps a payment step. I can understand choosing which services to connect, because the list resembles familiar application permissions: mail is allowed, a messaging app is not, and purchases require approval.
Meta already provides substantial controls of this kind. Its Muse help documentation describes connector settings, approval options ranging from one-time permission to broader defaults, an activity log, and the ability to disconnect a service. The company says users explicitly grant access and should review those permissions. These are real safeguards, even if they do not establish that every route into a particular application’s content is covered.
The difficulty begins when the operating system offers a different route. Android’s notification-listener interface allows a suitably authorized application to receive notifications from other applications. Its purpose is wider than AI, and a user must grant the relevant access, but a broad permission can be repurposed to read text from an app the user never connected to an assistant. On a desktop, accessibility capabilities and screenshots can likewise reveal information the agent encounters while controlling the device.
Those permissions have legitimate uses. Notification tools, assistive technology and automation software existed long before contemporary agents, while preventing their use indiscriminately would harm people who rely on them. Yet a permission granted for one purpose does not communicate every purpose the user is willing to accept. The ordinary interpretation of “I did not connect Signal” is narrower than “some other permission might still expose Signal’s visible content.”
Reuters also reported employee complaints about unsolicited messages, an unauthorized booking and bank details that appeared to belong to other users. Meta described the examples as isolated, said some apparent account mix-ups were hallucinations and was addressing identified issues. Reuters could not establish whether the unfamiliar financial information came from another person or was invented. Those unresolved incidents should not be turned into confirmed cross-account leaks; the Signal access reproduced with independent testers is sufficient to establish the permission-boundary question.
On November 16, two days before the Reuters investigation appeared, Apple announced changes in macOS 27.3 beta 2. It split a broad control into System & App Protection Override and Accessibility, introduced more granular choices involving particular apps or folders, and added options to ask each time and show an indicator when one app controls another through the relevant override.
These are material improvements, not a hypothetical remedy. Apple’s announcement also says previously granted app permissions remain in effect and that developers must prepare their products to function with narrower access. Because the changes were introduced in a developer beta, they should not be described as protection already present on every released Mac, let alone an equivalent policy operating across Android, Windows and every agent.
The product gap has therefore narrowed without disappearing. Better permission prompts help users understand broad access when they grant it. A stronger negative rule would let them decide that a particular application must remain excluded even when another allowed capability could expose its content. That requires coordination between the operating system, applications and agent providers rather than an extra warning in the agent’s chat window.
Signal’s president, Meredith Whittaker, made a similar request in the Reuters reporting: application developers should have consistent technical ways to mark their apps off-limits to AI agents, and operating systems should enforce those boundaries. Her proposal introduces another legitimate interest. A message belongs to a conversation between people, and the recipient granting an agent broad device access does not necessarily mean the sender expected that agent to read their words.
There is an unavoidable tension here. An absolute publisher-controlled ban could interfere with approved accessibility or personal automation, while leaving the choice entirely to an agent developer would put an important boundary in the hands of the software being restricted. A workable design needs to distinguish assistive access from autonomous agent use, make exceptions intentional and use enforcement points that do not depend on a model remembering instructions.
Imagine opening device settings and marking Signal: off-limits to AI agents. The ordinary expectation is simple: the travel assistant can read my permitted calendar and booking messages, but it cannot read Signal conversations merely because one of them happens to be visible elsewhere. For the setting to be meaningful, its scope has to survive more than one integration screen.
Connector access. An agent should not be able to connect directly to the excluded application or reuse an existing connector to retrieve its content. Product-level controls already address parts of this case.
Notifications. Previews originating from the protected application should not become readable input to an agent through a broadly authorized notification listener. Denying only the direct connector would leave the route used in Reuters’ Android reproduction untouched.
Screen and accessibility access. An agent controlling another window should not acquire the protected application’s text through screenshots, interface inspection or accessibility output. This is difficult when information from several apps appears on one desktop, so the system may need redaction or a refusal rather than pretending it can always separate the pixels perfectly.
Exceptions and audit. If the user intentionally wants a one-time interaction, the system should explain the specific application and requested access, then record the decision. An instruction such as “please ignore Signal” inside a chat is not a substitute for enforcement outside the model.
These are requirements for a proposed experience, not a claim that one current consumer operating system implements them consistently. Some content will also have been voluntarily copied into another permitted document or exported to a different device. A rule about access to the Signal application cannot guarantee that every quotation of a Signal message disappears from the world, and the interface should say where its authority ends.
A useful test would be straightforward to describe even if implementation is hard: mark the application off-limits, try a direct connector request, then try a notification, a visible conversation window and an accessibility-driven action. The agent should be able to explain why it cannot complete those requests, while unrelated calendar or travel tasks continue working. A permission change should be tested again after an application or agent update, because new capabilities can create new routes.
One tempting answer is to instruct the assistant never to read private messages. That may be helpful for ordinary task guidance, although a language instruction is a fragile place to enforce a security decision when the agent has tools capable of reading what it was asked to avoid. The operating system knows which application produced a notification or owns a window; an assistant interpreting the resulting text may no longer have that reliable origin information.
The most credible design therefore attaches restrictions to access mediation and source identity. The operating system can deny or filter protected streams before they reach the agent, while providers must respect the restriction across their connectors and tools. Users should be able to inspect the effective policy, including exceptions and old permissions that still apply. If an agent needs more authority to complete a particular task, its request can fail visibly rather than silently using another route.
This is also why a ban on all device automation would miss the user demand. People want assistants precisely because they can work across services, and some tasks genuinely benefit from seeing what’s on the screen. The better distinction is between the work the person authorized and information that remains outside that assignment, even when the operating system technically makes it reachable.
Connector and approval controls. Muse and other agent products already let users select connected services, review actions and revoke supported permissions.
Operating-system permission gates. Android notification access, macOS accessibility and screen-related permissions have documented authorization and management mechanisms.
More granular Mac controls. Apple’s November 16 developer beta adds app- or folder-scoped choices, repeated permission prompts and visibility into certain cross-app controls.
Security boundaries and audit mechanisms. Operating systems can mediate access to named resources, while agent products can log what tools and permissions they used.
One understandable negative rule. A user should be able to exclude a particular application from autonomous agents without translating the decision into every separate connector and permission surface.
Cross-route enforcement. Notifications, screenshots, accessibility output and direct integrations need consistent treatment of a protected application’s content.
Transparent exceptions. Assistive access and explicitly authorized one-time tasks need well-defined handling rather than a blanket loss of accessibility or a silent bypass.
Verifiable behavior across updates and platforms. A prohibition should show which routes it covers, preserve meaningful restrictions when capabilities change, and admit where enforcement is incomplete.
The inventory does not imply that these controls are technically impossible or completely absent. Apple is already moving in this direction. What is not yet established is a consistent, comprehensible experience that follows the protected application across the different ways an agent can encounter its content.
The Signal example matters because there was nothing exotic about the underlying access. A user can approve notification or accessibility permissions for a legitimate reason, then discover that a general-purpose agent can use those capabilities to infer something about a private conversation. That is a problem with the meaning and composition of permissions, even when no encryption fails and no unauthorized operating-system exploit occurs.
An agent that books travel, organizes files or follows up on work should be able to succeed with deliberately limited access. Where its permissions stop ought to be determined by controls a person can inspect, rather than by whether the agent happens to discover a convenient alternative path through the device.
I want to be able to mark an app off-limits once, understand the exceptions, and trust that the next helpful feature will not quietly redefine the rule. The more capable personal agents become, the less reasonable it is to make every user audit an entire operating system before granting one assistant a useful task.
Thirstday #12: Why Can’t I Make an App Off-Limits to Every AI Agent? was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.