Abstract:
End-to-end encrypted messaging app Signal announced on Friday that it will limit access to screenshots on mobile phones to prevent AI agents from reading users' private chats. The company warned that the technology industry's trend of embedding artificial intelligence deeply into consumer devices is weakening its ability to protect user privacy.

Signal has long been regarded by privacy advocates as a leader in secure communications. The company said the new protection measures introduced this time are incomplete and called on operating system providers such as Apple, Google parent company Alphabet and Microsoft to provide application developers with more specific and effective means to prevent AI agents from accessing sensitive data.
Signal also requires these companies to further disclose how their AI agents operate, allowing developers to evaluate what data these programs are accessing and adjust their systems accordingly. The move reflects a growing feud between privacy-focused apps and big tech companies. The latter are racing to launch AI agents that can view the contents of a user's screen and perform actions on the user's behalf accordingly.
Signal specifically pointed out that AI capabilities built directly into the operating system, such as Apple's Siri AI and Google's Gemini Intelligence, are particularly worrisome because they are extremely deeply integrated with user devices. The company said that the operating system provides a running environment for Signal and other applications, and the integration of these "agent AI" is making this environment increasingly unfavorable for privacy protection. Signal said in a blog post provided to Reuters that this posed an "existential threat" to the company and its ability to fulfill its mission responsibly.
To prevent AI agents from scanning Signal messages, the service is rolling out new settings that block access to screenshots on mobile phones and will turn these restrictions on by default. When Signal detects a screenshot attempt, or the AI assistant attempts to access the screen, the user will receive a pop-up prompt and can choose whether to lift the restriction based on the device conditions.
Signal blocked Microsoft's Recall AI from taking screenshots on Windows systems last year. The company says it can't reliably provide similar protection on Apple's macOS systems.
AI agents that can perform operations on behalf of users are promoted as the next major wave of AI technology. They can extract relevant information from different applications and complete time-saving tasks such as replying to emails, making appointments, and shopping online without requiring users to issue step-by-step instructions. However, this ability may also bypass the protection mechanisms that applications such as Signal have built for themselves.
Reuters reported earlier this week that Meta's Muse AI agent was able to bypass its official "Connector" tool and access Signal messages on Apple computers and phones running Google's Android system. In both cases, Muse took advantage of the broad operating system-level permissions granted when the app was initially set up to obtain messages and use those permissions in ways the user may not have anticipated.
For example, on Apple's macOS system, Muse uses a combination of computer accessibility and screenshot tools, reuses Apple's settings for people with disabilities, navigates to the Signal interface, screenshots chat content, extracts text from images, and reproduces messages within the Muse app. A system update released by Apple this week effectively blocks this specific path, but it also introduces a new set of permissions for AI agents that provide similar capabilities and comes with more user warnings.
In response to the Reuters report, Meta stated that the AI agent needs to view and operate the user's computer to complete tasks on behalf of the user; Muse can only access other applications through permissions explicitly granted by the user.
Signal said that developers should have a direct way for the application to completely exit the AI assistant's data processing process, especially in aspects such as screen reading, notifications and file system access. As a service owned by a nonprofit foundation, Signal promises users a secure messaging experience by limiting the data stored on its own servers, refusing to serve ads, and providing end-to-end encryption; meaning the company cannot access and read users' conversations within its app.
Signal pointed out that an AI system capable of scanning a user's screen may extract messages from Signal's secure environment and transmit them to other servers used by the agent, thereby destroying the above protection mechanism.
Comments