Back to Inside Anthropic’s IPO + How AI agents spy on your messages, with Meredith Whittaker.
How Can an AI Agent Read Signal Messages?
Meredith Whittaker says an AI agent on a phone or computer can erase the privacy Signal is able to offer. Encryption still hides a message on the way. It does not help if the other person's agent is allowed to read the screen, copy what is on the device, or pull the text out of Signal to write a summary.
She describes the steps as ordinary plumbing: a screenshot, storage, character recognition, then a copy sent to a remote server for inference. That can expose the person who sent the message, not only the person who turned the agent on. A reporter was warned not to leave Signal open on a desktop that also runs an agent. Assistants such as Muse, advertised around San Francisco, can build dossiers of contacts and of how you write to them. If that design stays in place, Signal cannot keep working with its integrity intact.
Inside Anthropic’s IPO + How AI agents spy on your messages ·
From the transcript
And then also think about this in, you know, the kind of plumbing and guts and material reality of like, you know, where is the server infrastructure? You know, is it screenshotting your device using some accessibility hook in order to send those screenshots into a database in order to run inference so that it can sort of come back with a suggestion date for your birthday party and do its agentic magic as your shiny little helper.
And so I do think it's not that people are less privacy conscious in terms of their values, in terms of not wanting that database of everything they've said out in the world. It's that there is a very powerful narrative that is making it very difficult to remember these fundamentals and why that is and what type of inoculate we're dealing with is maybe beyond the scope of my psychological training.