Google DeepMind researchers used an AI framework called AlphaProof Code to mathematically check key cryptographic guarantees in parts of the Signal protocol. The work adds an independent check of software that protects communications on Signal, Google Messages and WhatsApp, while offering an early demonstration of how AI might help with formal verification.
A focused check of the Signal protocol
Formal verification uses mathematics to test whether software meets specified properties. In security work, that can help establish whether code behaves as intended and supports particular protections. The approach has long been seen by its proponents as a powerful way to reduce the risk of critical vulnerabilities.
DeepMind’s project took four weeks and was conducted as a proof of concept. It examined key cryptographic guarantees in the Signal protocol, but it did not formally verify the entire Signal mobile app. The researchers describe the result as one step in a much larger process.
The analysis did not discover new vulnerabilities in the portions it assessed. The team also tried the framework on older Signal versions with known, long-fixed vulnerabilities. AlphaProof Code identified those issues and showed that the protocol could not be formally verified in those versions.
AI can help with a difficult process
Formal verification requires people to define the software properties they want to prove, express them as formulas and develop the proof. Those steps can take substantial time. Agentic AI may speed up parts of that work by helping produce and check proofs, making analyses like DeepMind’s more practical to carry out.
The researchers say they built a process in which AI agents can contribute and correct their work at many stages, with people providing input and review at important points. The aim is collaboration between people and AI, rather than having a system set its own goals and judge its own results.
That distinction matters because a proof is only useful if the properties being tested are meaningful and the verification process is trustworthy. Human review remains part of the approach described by the researchers.
What the result does and does not establish
The work provides an additional mathematical evaluation of selected parts of the Signal protocol. Signal’s chief technology officer, Ehren Kret, said in a statement that the broader approach could expand on earlier verification work that covered smaller portions of the code, and that rigorously checking the implementation gives the organization greater confidence in its privacy guarantees.
DeepMind’s researchers had a head start from past formal verification work conducted by the Signal Foundation. They are releasing the specification and proof they developed, which may help others working on formal verification.
Still, a successful check of selected protocol properties is not a claim that every part of the app is free from flaws. The researchers’ own description leaves clear limits: the mobile app as a whole was not formally verified, and the project is an early milestone rather than a comprehensive security guarantee.
Scaling and cost remain unanswered
For services used in high-stakes settings, verifying every code base and software supply chain would offer strong assurances. Google Cloud and DeepMind research vice president Pushmeet Kohli described that possibility in terms of organizations being able to prove that implementations meet desired properties.
Whether AI-assisted verification can reach that scale at a workable cost is unresolved. DeepMind provided the Signal analysis free of charge as a research project, and Google declined to disclose its cost. The company also declined to say whether it plans to offer a commercial formal verification product or service.
For now, the project shows that AI can contribute to a rigorous review of an important cryptographic protocol, and can detect known flaws in older versions. It does not show that the same method is ready to verify all software. The next challenge is to make the process useful across larger systems while preserving careful specification, human review and reliable proof.