VibeTimes
#기술

What Signal Couldn't Block: AI's Access to the File System

모민철모민철 기자· 11/22/2026, 6:18:07 AM· Updated 11/22/2026, 6:18:07 AM

Encrypted messenger Signal was unable to stop Meta's AI assistant 'Muse' from reading conversation content and displaying notifications. The case shows that encryption alone may not protect private conversations. Wikitree confirmed the case through reporting. One Meta employee received a notification from Muse urging them to reply to a message from their sister — a message that had arrived via Signal. In an internal post made public earlier this week, the employee wrote that it "seems like an extreme privacy violation," adding, "I don't remember granting access to other apps, or ever being asked for permission."

Signal, together with the AI Now Institute, reproduced the same situation on an Android phone and a Mac. On Android, Muse read the full text of messages that appeared in the notification shade. On the Mac, it used accessibility tools and the screenshot function to get into the Signal app, capture the screen, and extract text from the image, converting it back into text. Both experiments relied on permissions users had granted during initial setup, and Muse gained access without going through the official integration tool, the 'Connector'. Meta said Muse's paths into other apps are limited to permissions explicitly granted by users.

End-to-end encryption only protects a message while it is in transit. Once a message arrives, it must appear on screen through a notification or the chat window, and if AI can access that, it can read the same content. According to Signal's blog post, Signal has no say over whether an assistant processes the text it captures on the device or sends it to a remote server. As a result, the privacy of everyone in the conversation could be at risk, Signal explained.

Meredith Whittaker, president of the Signal Foundation, traced the root of the problem to accessibility settings. Accessibility features, originally designed to help people with disabilities, are being used by agents as "a master key that opens the door to AI agents," she noted. She said more tools are needed to prevent third-party agents from accessing app data with sweeping permissions.

Admitting its limits, Signal's defenses vary by platform. On iPhones running iOS 27, a 'Block Siri AI' setting is on by default, and on Android, 'Screen Security' turns on automatically when it detects an assistant attempting to read the screen. Windows 11 already has a similar setting introduced after the privacy controversy over Microsoft's 'Recall' feature in May. Signal warned that turning off these options could expose a user's content — and that of their conversation partners — to AI agents. At the same time, blocking screenshots alone cannot stop "more covert methods" that AI can exploit. A prime example is the route that pulls content stored on the device in through the file system (the repository of data stored on a device). Signal said it is "doing everything it can," but this route remains unsolved.

Signal called for "two common-sense fixes": a way for apps to reject AI data access and processing outright, and thorough documentation and technical specifications on how AI agents access data. Whittaker demanded that Apple, Google, and Microsoft disclose what their assistants can read and provide developers with means to block them.

The big tech companies have offered some user-side controls. Apple added an 'Excluded Apps' option in iOS 27.2 late last month to choose which apps Siri, Search, and Suggestions can access, and Google Gemini connections can be severed in its 'Connected Apps' settings. But what Signal wants is different in kind: a means for app developers to remove their own apps from AI's reach.

Related Articles