Formal verification has been the security world’s holy grail for decades, and AI may finally be making it less miserable. Google DeepMind researchers used an agentic framework called AlphaProof Code to mathematically prove several of the Signal protocol’s key cryptographic guarantees, and they did it in four weeks.
According to WIRED, the work is a proof of concept with two payoffs. It gives independent validation of the cryptography behind Signal, Google Messages and WhatsApp. And it shows what DeepMind’s verification tooling can do. The team found no new vulnerabilities in the portions of the protocol it examined. But when it pointed AlphaProof Code at older Signal versions with known, long-fixed flaws, the tool caught them and showed those versions could not be verified. That matters, because a verifier that only says yes is not much of a verifier.
The reason this is hard without AI is simple. Humans have to describe the properties they want proven, turn those descriptions into formulas, then write the proof. The last step eats the most time, and that’s the part agents can speed up. Still, the scope here was limited. DeepMind did not verify the entire Signal mobile app, and the team had a head start from earlier verification work by the Signal Foundation.
Signal’s CTO, Ehren Kret, said the approach allows much broader coverage than prior efforts. Pushmeet Kohli, Google’s Cloud chief scientist and DeepMind’s VP of research, pitched the bigger picture: banks and other firms could one day say they can prove specific properties of their code. Swarat Chaudhuri was blunter about the gap, saying there is a long way to go before this reaches “all the software of the world.”
That gap is the real story. Formal verification has existed for years in tools like Coq, Lean and Isabelle, and startups and labs are now racing to pair them with language models. DeepMind’s earlier AlphaProof work on math problems fed directly into this. But a proof of Signal, a well-studied protocol with prior verification behind it, is the friendly case. Messy enterprise code with unclear specifications is not.
Cost is the other unknown. DeepMind ran the analysis free as research, and Google declined to share what it cost or say whether a commercial verification product is planned. Without those answers, developers can’t judge whether this beats conventional audits and fuzzing on price.
One design choice deserves credit. Chaudhuri says the process keeps humans reviewing at key points, rather than letting AI write both the goals and the proofs, which would be like grading your own test. That middle ground is what makes the results believable.
So is this noise? No, but it’s early. For security teams, the release of the Signal specification and proof gives others something concrete to build on. Watch for pricing and a product before getting excited about your own codebase.
