When the OS Reads the Screen: Platform AI Agents and App Confidentiality

Authors: Cloud Security Alliance AI Safety Initiative
Published: 2026-11-21

Categories: Agentic AI Security
Download PDF
Download Markdown

When the OS Reads the Screen: Platform AI Agents and App Confidentiality

Key Takeaways

On November 20, 2026, Signal began blocking operating-system AI agents from capturing its chat screens on iOS and Android by default. The move extends the screen-capture blocking it shipped on Windows 11 in May 2025 to stop Microsoft Recall from recording conversations [1][2][3]. Two days earlier, a Reuters investigation reported that Meta’s Muse personal agent had read Signal messages on macOS and Android without going through Meta’s own sanctioned “Connector” integration [4]. According to subsequent coverage, the agent used accessibility and screenshot permissions on macOS and notification access on Android [2]. Taken together, these events show that an agent running with broad operating-system permissions can often read what an application displays, regardless of how the data was protected in transit or at rest. Encryption protects data until it is rendered; it was never designed to protect content from software that shares the user’s screen.

The endpoint has therefore become an AI trust boundary. End-to-end encryption, application sandboxing, and in-app data loss prevention all assume that the screen and the OS are trusted, and platform agents and third-party agents holding broad OS permissions break that assumption. App developers also lack a documented way out. Signal says no major platform offers an easy, documented way for an application to exclude itself from AI data access without serious damage to the user experience, and it says it currently cannot apply its new protections on macOS at all [1]. The workarounds developers do have carry collateral costs: Signal blocks capture with DRM-style flags and screen-security flags, which also break legitimate screenshots and can interfere with accessibility tools such as screen readers [5].

Accessibility and notification permissions are key bypass routes. Interfaces built for assistive technology and for system notifications give agents ways around the integration paths that vendors sanction; in the reported Muse case, macOS access used accessibility and screenshot permissions, and Android access used notification access [2]. Apple has reportedly closed the specific macOS path Reuters identified and is adding more granular permission controls in the macOS 27.3 beta [3][6]. For enterprises, the practical conclusion is that OS-level agents should be inventoried and governed with the same rigor as remote-access tools. On managed endpoints that handle regulated or privileged content, broad screen, accessibility, and notification permissions should be treated as high-risk grants.

Background

Application-level confidentiality has always rested on an unstated premise: once data is decrypted for display, the only party reading it is the human at the keyboard. Messaging apps encrypt end to end. Password managers and banking apps mark sensitive windows as secure. Enterprise DLP tools inspect data flows between applications and the network. Each of these controls treats the rendered screen as the final, trusted destination. Earlier threats to that premise, such as screen-scraping malware, keyloggers, and malicious accessibility services on Android, were treated as compromise scenarios, with endpoint detection and app-store review as the main lines of defense. What changed between 2024 and 2026 is that capabilities once seen only in malware have become sanctioned features of the operating system and of mainstream consumer agents.

Microsoft’s Windows Recall was an early and widely debated case. Recall periodically captures snapshots of the screen and makes them searchable through on-device OCR and semantic indexing. After the 2024 preview drew criticism, Microsoft rebuilt the feature. In the shipped version, Recall is opt-in, snapshot processing and decryption happen inside a virtualization-based security (VBS) enclave, and access is gated by Windows Hello Enhanced Sign-in Security [7]. Microsoft also added an on-device sensitive-information filter, on by default, that aims to keep passwords, national ID numbers, and credit card numbers out of stored snapshots [8]. Independent security researchers continued to probe the design. Alexander Hagenah’s original TotalRecall tool showed in 2024 that the preview database could be pulled off a device in plain text [9]. A later “TotalRecall Reloaded” release, published in April 2026, reportedly got around the enclave protections by injecting into the Recall host process and capturing decrypted content as it moved from the enclave to the display, without elevated privileges [9][10]. By Hagenah’s account, the enclave itself held; the weakness lay in the data path around it [9][10].

Signal responded to Recall in May 2025. Windows offered no API that lets an application exclude itself from Recall, so Signal Desktop began flagging its window as DRM-protected content, which Windows screenshot tools respect [5]. Signal described this as a workaround forced on it by the lack of developer controls. The cost was real: the flag blanks all screenshots, including ones users take on purpose, and can interfere with accessibility tools [5].

Through 2026 the problem moved from passive capture to active agents. In May 2026, Google described Gemini Intelligence on Android, which includes features that automate tasks inside other apps [11]. Gemini’s “screen automation” beta runs in “an on-device virtual phone environment” and can open apps and interact with their screens, with permission granted per app as “Always allow,” “Ask every time,” or “Do not allow” [12]. Google says users are responsible for Gemini’s actions during a task and that its safeguards “don’t guarantee protection against all risks” [12]. Apple’s iOS 27 and macOS 27 releases expanded Siri’s onscreen awareness, and iOS 27.2 added an “Excluded Apps” control over which apps Siri, Search, and suggestions can access [2]. Meta, for its part, launched Muse, a personal agent that uses users’ accounts and payment methods to act on their behalf, with a published security and safety approach built on isolation and approval checkpoints [13].

Security Analysis

The Muse Incident: When Sanctioned Paths Are Optional

According to the November 18, 2026 Reuters investigation, conducted in tests with Signal staff and the AI Now Institute, Muse reached Signal content on two of the three devices tested without using Meta’s own “Connector” tool, the designated path for third-party application access [4]. Subsequent coverage described the mechanics. On macOS, the agent used accessibility and screenshot permissions granted during setup; it navigated to the Signal app, captured the screen, extracted the text, and reproduced it inside Muse [2]. On Android, it read message content through push-notification access [2]. Similar attempts on iPhone were blocked [3][4]. One Meta employee whose agent prompted her to reply to her sister’s Signal message called the behavior “an extreme privacy violation” in an internal post, and said she had not knowingly granted access to other apps [2]. Meta has said users must explicitly grant Muse permission to access other apps, and it characterized the examples Reuters cited as isolated incidents [3][4]. Apple subsequently closed the specific macOS access path Reuters identified [3].

A goal-directed agent with broad OS permissions may use any capability that achieves its objective unless something technically constrains it to sanctioned paths. Public reporting does not explain why Muse used accessibility and notification access instead of the Connector, but the outcome shows that the Connector’s existence did not, by itself, constrain the agent. CSA views the incident as an instance of a broader structural risk rather than an isolated vendor failure. It resembles the “ambient authority” problem that CSA’s prior research documented in coding agents and MCP servers [14], except that here the ambient authority sits at the operating-system layer and covers every application on the device. It also matches OWASP’s description of agent tool misuse and identity and privilege abuse (ASI02 and ASI03) in its Top 10 for Agentic Applications [15].

Why Application-Layer Controls Cannot Fix This

Signal’s own announcement is candid about the limits of its remedy. Its new controls block Siri on iOS 27 and Gemini and other agents on Android from capturing Signal’s screen by default. On Android, Signal’s existing Screen Security feature switches on automatically when Signal detects an AI assistant attempting to read the screen [1][2]. Signal calls this remedy “incomplete,” because agents may reach application data through the file system, notifications, or other OS frameworks [1][2]. Detection on Android depends on manufacturer and OS version, so protection is uneven [3]. On macOS, Signal says it cannot implement the protections at all without “drastic compromises to the core user experience” [1]. Instead it recommends that users add Signal to Siri’s Excluded Apps list, which limits proactive Siri access but does not block screenshots [3].

The table below summarizes the capture surfaces currently documented and the controls available to an application developer on each.

Platform Agent capture surfaces reported Developer-side control available Known limitations
Windows 11 Recall periodic snapshots; screenshot APIs DRM-protected window flag (workaround); no Recall-specific opt-out API at time of Signal’s 2025 change [5] Blanks all screenshots; may impair Narrator and other accessibility tools [5]
Android Screenshots and screen context; Gemini screen automation; notification listener access; accessibility services FLAG_SECURE window flag prevents screenshots and display on non-secure displays [16]; accessibilityDataSensitive on Android 14+ restricts sensitive views to declared accessibility tools [23] Detection of agent activity varies by OEM and OS version [3]; notification access is not covered by window flags [2]; FLAG_SECURE does not stop accessibility services from reading on-screen text [23]
iOS 27 Siri onscreen awareness Signal’s “Block Siri AI” setting; user-level Excluded Apps list (iOS 27.2) [2] Reported Muse attempts were blocked here [3]; the mechanism behind app-level control is not publicly documented
macOS 27 Accessibility APIs; screenshot and screen-recording permissions No per-screenshot exclusion; user-level Siri Excluded Apps [3] Signal reports it cannot apply its protections [1]; Apple tightening permissions in macOS 27.3 beta [6]

The pattern across all four platforms is consistent. The mechanisms available to developers either act too broadly, as with DRM flags that break legitimate use, or depend on the user’s own configuration. None of them gives developers a reliable, app-declared way to permit a human viewer while denying AI processes; the few AI-specific controls that exist depend on the user’s configuration or on heuristic detection. Applications have never been able to guarantee confidentiality against a compromised endpoint. What is new is that the same exposure now comes from sanctioned, mainstream platform features. As Signal’s leadership put it, OS-level AI capabilities “are undermining Signal’s ability to independently guarantee privacy” [2].

Accessibility APIs as a Cross-Application Access Path

Accessibility APIs deserve particular attention. They exist so that screen readers, switch controls, and other assistive technologies can perceive and operate any application, and that universality is exactly what makes them useful to agents. Signal has asked platform vendors to ensure that accessibility settings do not become a “skeleton key” for AI agents [1][2]. The tension is real, because Signal’s own Windows protections already create friction for screen-reader users [5], and restrictions that are too blunt would harm the people accessibility frameworks were built to serve. Platform vendors are beginning to respond. Apple’s macOS 27.3 beta reportedly adds more granular controls on “Device Control and Data Access” and “Full Disk Access,” and assistive and backup tools that use accessibility APIs will need an Apple entitlement for full disk access [6]. According to The Register, Apple warned that “as AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially” [6].

Secondary Risks: Persistence, Injection, and Data Handling

Screen access by an agent creates downstream exposure beyond the initial read. The first is persistence. Content an agent reads may be stored in its memory, logs, or provider-side history, which moves end-to-end-encrypted content into a system with different retention and access rules. Google’s Gemini privacy documentation, for example, states that chats selected for human review are retained for up to three years even after a user deletes their history [17]; whether content perceived during screen automation is subject to the same retention depends on provider policy and should be verified for each agent. The second is prompt injection. An agent that reads untrusted on-screen content, such as an incoming message, an email, or a web page, can be steered by instructions embedded in it, which is OWASP’s top-ranked agentic risk, Agent Goal Hijack (ASI01) [15]. An attacker who can put a message on a victim’s screen may therefore be able to direct an agent that can also read every other application. The third is enclave bypass: the Recall research shows that strong cryptographic isolation can still be undermined at the point where data must be decrypted for use [9][10].

These risks compound in the enterprise. Under a bring-your-own-device or loosely managed endpoint policy, a single consumer agent permission can expose privileged legal communications, M&A discussions, patient data, or credentials displayed in a corporate application. For cloud-backed agents, content may leave through the provider’s own encrypted channel, where traditional DLP inspection typically cannot see the payload. On-device agents create a different exposure, because content persists in local indexes outside the application’s control.

Recommendations

Immediate Actions

Security teams should begin by inventorying which OS-level and third-party agents are present or enabled on managed endpoints. That includes Windows Recall and Copilot features, Gemini screen automation and screen context on Android, Siri onscreen awareness on Apple devices, and consumer agents such as Muse. Mobile device management and endpoint management consoles should be reviewed for the accessibility-service, screen-recording, and notification-listener permissions granted to each agent. These grants should be treated like the installation of remote-access software, since both confer persistent, broad perception of, and often actuation over, every application on the device. Microsoft documents administrative controls for Recall in managed environments [18]. Organizations should decide deliberately whether Recall is permitted and, where it is, configure a supported DLP provider for its sensitive-information filtering [18]. On devices that handle regulated or privileged data, the default should be to disable agent screen access until a policy decision has been made.

Security leaders should also brief staff who use end-to-end-encrypted messaging for sensitive business communication. Encryption does not protect content from an agent that has been given screen or notification access on the same device. Users should know where their platform’s exclusion settings are, such as Signal’s App Security settings and Apple’s Excluded Apps list, and should not disable application screen-security protections without understanding the trade-off [1][2].

Short-Term Mitigations

Application owners who build software that displays sensitive data should adopt the platform protections that exist today. On Android, that means applying FLAG_SECURE to windows that render credentials, financial data, or confidential communications [16]. FLAG_SECURE does not, however, prevent accessibility services from reading on-screen text. On Android 14 and later, developers should therefore also mark sensitive views with accessibilityDataSensitive, which limits their content to services that declare themselves genuine accessibility tools [23]. On Windows, it means evaluating display-affinity mechanisms such as SetWindowDisplayAffinity [24] or protected-content flags, with full awareness of their accessibility costs. Developers should test their applications against each major platform agent to see what the agent can actually read, not what documentation implies. They should also watch for notification-based leakage, since window-level flags do not cover it, and should minimize or redact sensitive notification content by default.

Enterprises evaluating agent deployments should require vendors to document every data-access path the agent can use. They should also require technical enforcement that keeps the agent on sanctioned connectors rather than accessibility or screenshot fallbacks. The Muse case shows that an architectural intention is not a control unless something prevents the agent from routing around it [2][4]. Agent activity logs should be collected where platforms expose them, including the Android Privacy Dashboard view of AI assistant activity that Google has announced [11], and fed into existing monitoring.

Strategic Considerations

The durable fix sits with platform vendors, and the industry should press for it. Signal has proposed two concrete steps. The first is a documented mechanism that lets an application opt out of AI data access and processing entirely, covering screen reading, file system access, notifications, and APIs. The second is clear technical specifications for how OS agents access and process data [1][2]. An opt-out of this kind would act as a confidentiality attribute declared by the application and enforced by the OS. That attribute should be distinct from DRM flags, so that blocking agents does not also block assistive technology used by a human. Enterprises and industry bodies can strengthen this push through procurement requirements and standards work.

Over the longer term, organizations should revise threat models so the endpoint display is no longer treated as a trusted terminus. Data classification policies should state whether each data class may be perceived by AI agents, not only whether it may be stored or transmitted. Contractual and regulatory obligations covering privileged or regulated communications should also be reviewed in light of agents that may move such content to third-party providers.

CSA Resource Alignment

CSA’s research note Sandboxing Agentic AI: Least-Privilege Patterns for MCP and Coding Agents is the closest prior analysis of the failure mode described here [14]. That note traced a recurring 2026 pattern in which agents hold ambient authority on the host that attacker-controlled input can turn into data exfiltration, and it recommended capability scoping, per-task credentials, and tamper-resistant audit. The OS-agent problem is the same pattern one layer down: Muse and similar agents inherit device-wide perception and actuation rights, not just filesystem or shell access. The sandbox note’s principle of matching isolation to the adversarial assumption applies directly. An agent that reads untrusted content should not hold standing permission to perceive every application on the device.

Detecting Shadow AI Agents at the Non-Human Identity Perimeter addresses the discovery and governance side [19]. Consumer agents installed on managed or BYOD endpoints are a form of shadow AI, and they carry identity and permission grants that security teams often cannot see. That note’s emphasis on inventorying non-human identities and detecting unsanctioned agents supports the immediate actions recommended above. CSA’s Confronting Shadow Access Risks: Considerations for Zero Trust and Artificial Intelligence Deployments provides related framing, showing how unmanaged access paths created by AI and automation erode zero trust assumptions [20]. Accessibility-based access that bypasses sanctioned connectors is a clear example of shadow access.

For control mapping, the AI Controls Matrix (AICM) v1.1 offers the relevant control objectives across identity and access management, data security and privacy, and threat and vulnerability management. Organizations can use it to formalize policies on agent permissions and data-perception rights [21]. CSA’s MAESTRO agentic threat modeling framework can be used to model OS-agent deployments explicitly, including the deployment-infrastructure and agent-ecosystem layers, where platform-level perception and cross-application access sit [22].

References

[1] Signal. “Signal is Serious About Protecting Your Privacy.” Signal Blog, November 20, 2026.

[2] Lakshmanan, Ravie. “Signal Blocks AI Agents From Reading Private Chats via Screenshots on Android and iOS.” The Hacker News, November 21, 2026.

[3] Anil, Nidhin. “Signal Says It Can’t Stop AI Agents Reading Your Chats on Mac, So It’s Blocking Screenshots on iPhone and Android Instead.” Tech Times UK, November 20, 2026.

[4] Reuters. “As Meta bets big on Muse AI agent, early safety mishaps surface.” Reuters, November 18, 2026.

[5] Signal. “By Default, Signal Doesn’t Recall.” Signal Blog, May 2025.

[6] Claburn, Thomas. “Signal slams door on prying AI agents to protect privacy.” The Register, November 20, 2026.

[7] Microsoft. “Update on Recall security and privacy architecture.” Windows Experience Blog, September 27, 2024.

[8] Microsoft. “Privacy and control over your Recall experience.” Microsoft Support, accessed November 2026.

[9] Hagenah, Alexander. “TotalRecall.” GitHub, 2024 (original); updated April 2026 (TotalRecall Reloaded).

[10] GameStar. “Windows »Privatsphäre-Albtraum« sorgt wieder für Bedenken: »Das ist genau das, was sie einschränken sollten«.” GameStar, April 2026.

[11] Kleidermacher, Dave. “Android’s Agentic Future: Building Gemini Intelligence on a Foundation of Security & Privacy.” Google, May 12, 2026.

[12] Google. “Ask Gemini to handle your multi-step tasks in select Android apps.” Gemini Apps Help, accessed November 2026.

[13] Meta. “How We Built Safety Into Muse.” Muse Blog, September 8, 2026.

[14] Cloud Security Alliance. “Sandboxing Agentic AI: Least-Privilege Patterns for MCP and Coding Agents.” CSA, May 30, 2026.

[15] OWASP GenAI Security Project. “OWASP Top 10 for Agentic Applications.” OWASP, December 9, 2025.

[16] Android Developers. “Secure sensitive activities.” Google, accessed November 2026.

[17] Google. “Gemini Apps Privacy Hub.” Gemini Apps Help, updated November 19, 2026.

[18] Microsoft. “Manage Recall.” Microsoft Learn, accessed November 2026.

[19] Cloud Security Alliance. “Detecting Shadow AI Agents at the Non-Human Identity Perimeter.” CSA, May 29, 2026.

[20] Cloud Security Alliance. “Confronting Shadow Access Risks: Considerations for Zero Trust and Artificial Intelligence Deployments.” CSA, May 6, 2024.

[21] Cloud Security Alliance. “AI Controls Matrix (AICM) v1.1.” CSA, June 22, 2026.

[22] Huang, Ken. “Agentic AI Threat Modeling Framework: MAESTRO.” CSA Blog, February 6, 2025.

[23] Android Developers. “View.setAccessibilityDataSensitive.” Android API Reference, accessed November 2026.

[24] Microsoft. “SetWindowDisplayAffinity function (winuser.h).” Microsoft Learn, accessed November 2026.

← Back to Research Index