Google DeepMind used agentic AI to mathematically prove key cryptographic guarantees of the Signal protocol, the encryption base for Signal, Google Messages and WhatsApp. The proof-of-concept took four weeks and relied on a new framework called AlphaProof Code. The result matters because it shows AI can perform formal verification that previously required long manual work.
Four-week Signal verification project
The analysis confirmed the properties DeepMind examined and found no new vulnerabilities in those parts of the protocol. As a control test, researchers ran AlphaProof Code on older Signal versions with known and long-fixed flaws. The system correctly flagged them as unverifiable, which confirmed that the method distinguishes secure code from insecure code. DeepMind is releasing the specification and proof to support further verification work.
The project followed a human-AI collaboration process in which agents contributed and self-corrected at almost every phase while humans reviewed key junctures. Research scientist Swarat Chaudhuri said the team did not want AI to generate goals and proofs against self-written tests. Agents therefore worked inside human-defined properties and checkpoints rather than operating autonomously. That design kept the proof tied to intended security guarantees instead of machine-invented criteria.
The team had a head start because the Signal Foundation had already conducted prior formal verification work. Earlier efforts covered smaller portions of the code, while DeepMind expanded the scope with AI assistance. Signal chief technology officer Ehren Kret said the approach enables significant expansion of verified code. He linked the work to growing AI-driven threats and stronger confidence in privacy guarantees.
What AI verification means for business
For companies that depend on messaging, payments or document exchange, verified cryptography lowers the risk of hidden flaws in a core trust layer. Banks and other regulated firms that need documented assurance could point to proven properties rather than test coverage alone. Large organizations may first apply such checks to authentication, key exchange and payment modules. Small firms benefit indirectly when widely used protocols and libraries carry stronger guarantees.
The method remains limited in scope and cost. DeepMind did not verify the entire Signal mobile app, and researchers described the result as one milestone in a longer process. The analysis was done free of charge as research, Google declined to share its cost, and it declined to comment on commercial product plans. Buyers should therefore ask vendors what code was covered, what properties were proved, and what human review supported the result.
A marker to watch is whether DeepMind or others repeat this level of verification on another widely used protocol within a comparable time frame. A second public release with specification, proof and independent review would show the process transfers beyond Signal. Without that repetition, the project stays a strong demonstration rather than a repeatable service.
