November 21, 2026AgentsAgent-OperableInfrastructure

Signal Treats the OS Agent as the Adversary

For a decade the threat model of an encrypted messenger was the network and the server. On Friday Signal added a new one: the operating system's own AI assistant. In a post on November 20, Signal said it will restrict agentic systems built into iOS 27 and Android from capturing what is on its screen, and called OS-level agent integration "an existential threat to privacy at the application layer". Post: https://signal.org/blog/protecting-your-privacy/

The mechanics are blunt because the tools are. On iOS there is a new Block Siri AI setting that prevents screenshots inside the app. On Android, Signal automatically turns on its Screen security setting when it detects an AI assistant trying to read the screen. Users who want screenshots back can turn it off, at the cost of exposing their contacts' messages to the agent. Signal did this once before, against Windows Recall last year. Two admissions stand out: the fix is incomplete, since it does nothing about agents reading app data from the filesystem, and it cannot be done at all on macOS, because Apple offers no reliable way to block screenshots there.

The real complaint is documentation. Signal says it spent months doing forensics on Android and iOS just to learn what these assistants actually access, and asks Apple, Google and Microsoft for two things: a well-documented way for apps to opt out of AI data access across screen reading, files, notifications and APIs, and proper specs on how the agents handle data. It also names third-party agents like Meta's Muse, which lean on accessibility permissions that were meant for people, not as a skeleton key for software.

The same day, a follow-up to Apple's agent plans (we covered Gurman's Siri agent report on November 8) shows the other side of the fence. Bloomberg spotted an Apple job listing for an agent that "browses the web, works across apps, runs commands, writes and debugs code, and calls services". Bloomberg says Apple has eyed shipping as soon as the end of next year but the work is early, and that macOS 27.3 will warn users against giving third-party agents full access. TechCrunch summary: https://techcrunch.com/2026/11/19/apple-may-be-developing-its-own-ai-agent-that-runs-your-mac/

Put those together and the shape of the next fight is clear. Platform owners are fencing out third-party agents while building their own, and app developers are fencing out the platform's agent. Everyone wants agents to act for users; nobody has agreed who gets to see what. The missing piece is not a smarter model. It is a permission API with an opt-out that apps can actually call.
← Previous
Coding Agents Misreport 37% of the Time, and Compaction Makes It Worse
Next β†’
Two Million Canaries: Your Cheap Inference Route Is Reading Your Prompts
← Back to all articles

Comments

Loading...
>_