Google's AI Proves That the Code Protecting Signal's Private Messages Does What It Promises
Google DeepMind, Google's AI research lab, says its AI programs wrote a step-by-step mathematical proof that the core code behind Signal, the private messaging app, keeps messages secret the way its designers intended. The work took four weeks; the company says it usually takes expert teams years. It covers the app's central locking code, not the whole app, and the results are Google's own.
Google DeepMind, Google's AI research lab, today announced that its AI programs have produced a mathematical proof that the core code behind Signal keeps its promises. Signal is a messaging app known for privacy, and the same underlying method, called the Signal Protocol, scrambles messages in other apps too, including Google Messages. Google describes it as the technology protecting billions of private conversations.
Code that scrambles messages is usually checked by people reading it and by testing it, which can miss mistakes. A mathematical proof is stronger: it shows, step by step, that the code must behave as intended in every case, not just the ones someone thought to try. Writing such proofs by hand is slow, painstaking work, which is why it is rarely done for large programs.
In this project, human experts first wrote down exactly what the code is supposed to guarantee, for example that messages stay protected even after someone briefly gets into a person's phone. Then Google's AI programs, built on its Gemini chatbot technology, wrote the proofs. They produced more than six million lines of them in Lean, a computer language that checks every logical step automatically, so an error in the AI's reasoning would be caught. The programs flagged about one percent of the code for a person to look at again. As a test, the team put back two old mistakes that Signal had already fixed, and the system proved that those flawed versions were not secure.
Google says the work took four weeks, a task that it says has historically taken expert teams years. Signal's chief technology officer, Ehren Kret, said: "Using AI agents to translate live code into valid Lean theorems is a major breakthrough." Google is publishing the proofs so that others can check them.
The result is narrower than it may sound. It covers selected guarantees of Signal's central locking code, not the whole app, and not other apps that use the same method. It assumes the underlying mathematics of scrambling is sound, so it says nothing about weaknesses there. A proof is only as good as the list of guarantees people wrote down at the start; if that list left something out, the proof would not catch it. And the timing, the one percent figure and the bug test are Google's own reports, with no outside review yet. Signal took part in the work, so its praise is not independent either.