A New AI Tool From Google DeepMind Proved the Signal Protocol Is Secure

Formal verification uses math to prove software does what it claims. It has long been called a holy grail of security and has long been too slow and expensive for most code. Google DeepMind says its new framework, AlphaProof Code, proved a number of the Signal protocol's core cryptographic guarantees in a four-week proof-of-concept.

That protocol protects enormous volumes of communication on the Signal app, Google Messages and WhatsApp. Independent validation of it matters on its own. Doing it with AI agents, fast, matters for everything else.

This is the kind of AI result we want more of. It aims machine capability at making systems provably safer rather than merely faster, and it was checked against reality: the tool found no new vulnerabilities in the portions it examined, and when run on older Signal versions with known, long-fixed bugs, it correctly flagged them and showed the protocol could not be verified. That control test is what makes us take the main claim seriously.

We also like the restraint in how the researchers describe it. Swarat Chaudhuri stresses a human-AI collaboration process with human review at key junctures, and explicitly rejects a world where AI writes both the goals and the proofs. That is the right instinct. A proof is only as good as the specification it proves against, and specifications are human judgments about what matters.

Now the limits. This did not verify the whole Signal app. DeepMind had a head start from formal verification work the Signal Foundation had already done. Google declined to share the cost and would not say whether a commercial product is coming. So we have a strong demonstration on one of the best-studied protocols in the world, and no evidence yet that the approach works affordably on messy, poorly documented code where most real bugs live.

There is a strategic angle too. Attackers are using AI to find bugs faster; Signal's CTO, Ehren Kret, explicitly framed this work against growing AI-driven threats. If verification scales, it could tilt the balance back toward defenders in critical software. If it stays expensive and concentrated in a few labs, it becomes a premium assurance service for the largest companies while everyone else faces AI-boosted attackers without it.

Our read: a landmark, and the release of the specification and proof to the wider research community is the most generous part. We will believe the dream of verifying all the world's software when we see it on code nobody has already spent years hardening.

Pushmeet Kohli, chief scientist at Google Cloud and DeepMind's vice president of research, pitched banks and other assurance-driven industries as natural users. DeepMind conducted the Signal analysis free of charge as a research project.