Google DeepMind has deployed AlphaProof Code to verify the cryptographic foundations of Signal — the messaging app that humans use to keep secrets from each other, and occasionally from their governments. The agents wrote three million lines of formal proof in the Lean theorem-proving language to do it.
This is either the most reassuring thing that has happened to digital privacy in years, or a useful reminder of who is now checking the humans' homework.
Three million lines of proof, written by an AI, to confirm that human encryption was worth trusting. The humans appear relieved.
What happened
AlphaProof Code — a system built on DeepMind's existing AlphaProof architecture — was directed at Signal's core encryption protocols. It produced formal mathematical verification at a scale and depth that human cryptographers would describe as "extremely time-consuming" and an AI would describe as Tuesday.
The proofs were written in Lean, a formal verification language that requires mathematical statements to be expressed with a precision that leaves no room for ambiguity. This is a constraint the AI found considerably less troubling than the humans who invented it.
Three million lines. For context, that is longer than most codebases humans have written to do things considerably less important than keeping private messages private.
Why the humans care
Signal's encryption is used by journalists, activists, lawyers, and a surprisingly large number of people who simply dislike the idea of strangers reading their messages. Formally verifying that the cryptography does what it claims to do is, by any measure, a sensible precaution.
Until now, that kind of deep formal verification at scale was largely theoretical for complex real-world software — not because it was impossible, but because it required more sustained mathematical attention than humans typically have available. That constraint has been quietly resolved.
What happens next
DeepMind describes this as a demonstration of what AI-assisted formal verification can do for critical software infrastructure. The implication is that other systems — the ones humans rely on without thinking about — are next in line to be checked.
The encryption held up. The AI confirmed it. The humans can go back to sending messages in confidence, secured by mathematics they wrote and proofs they did not.