Google DeepMind's AlphaProof Code Verified Signal's Encryption in Four Weeks

Google DeepMind took four weeks to mathematically verify selected cryptographic guarantees of the Signal protocol — the encryption design behind Signal, Google Messages and WhatsApp — using a new AI framework called AlphaProof Code. The work, reported by WIRED, found no new vulnerabilities in the portions examined, and DeepMind is releasing the specification and proof it developed so others can build on them.

For security practitioners, that's the headline: a demonstration that AI-assisted formal verification can work on real-world, security-critical code rather than textbook examples. But the project came with clear scope limits and unanswered questions about cost and commercialization.

What was actually verified?

DeepMind researchers applied AlphaProof Code to "a number of the Signal protocol's key cryptographic guarantees." Formal verification uses mathematics to prove software meets specified requirements — in this case, that the protocol's cryptographic properties hold as designed.

The analysis found no new vulnerabilities in the examined portions. As a validation step, the team ran AlphaProof Code on older Signal versions containing known, long-fixed vulnerabilities. The system identified those flaws and showed the protocol could not be formally verified in those versions — a useful sanity check that the method catches real bugs, not just confirms clean code.

The researchers are now releasing the specification and proof they developed to support other formal verification work.

What does this change for developers?

Not full coverage, for starters. The project did not formally verify the entire Signal mobile app, nor every part of the services that use the protocol. DeepMind researchers emphasize the findings are "just one milestone in a larger process."

Swarat Chaudhuri, a DeepMind research scientist, told WIRED: "We are very excited by this result, but there is a long distance to go before we are able to scale this to all the software of the world, and we want to do that."

The four-week timeline also benefited from prior formal verification work by the Signal Foundation. The DeepMind team said that earlier work gave them "a head start to a degree." Signal chief technology officer Ehren Kret noted that "while prior formal verification focused on smaller portions of the code, Google DeepMind's approach enables significant expansion."

So if you're a security team hoping to point an AI at your codebase tomorrow, temper that expectation — the DeepMind result leaned on existing verified groundwork and human-defined requirements.

Human-AI collaboration, not autonomy

DeepMind describes the workflow as a "human-AI collaboration process": agents contribute and self-correct at almost every phase, while humans provide input and review at key junctures. People still define what properties to prove and review critical stages.

Chaudhuri pushed back on the idea of fully autonomous proving. He told WIRED the team does not want a "vibe world where AI is generating the goals and AI is generating the proofs and basically taking tests that are self-written," adding that "there's a middle ground."

The commercial question

DeepMind conducted the analysis free of charge as a research project. Google declined to share the cost of the work and declined to comment on whether it plans to release a commercial formal verification product or service.

Pushmeet Kohli, chief scientist at Google Cloud and DeepMind's vice president of research, framed the potential value broadly: "If you can now verify the correctness of these implementations, then you can rest easy... You can imagine banks, imagine all those other companies who rely on assurances and who worry about bugs who could now say, 'I can prove these properties.'"

That's the market signal worth watching. Banks, infrastructure operators and enterprise software vendors spend heavily on audits and penetration testing; provable correctness would be a categorically stronger assurance.

Scale and cost remain the open questions

The central constraint is whether AI-automated formal verification can cover far more software at a workable cost. Formal verification has historically required substantial human effort to describe intended behavior and build proofs. AlphaProof Code aims to shorten the proof-building phase, but the specification-writing and review steps still demand expertise.

DeepMind's ambition reaches well beyond messaging apps, yet Chaudhuri acknowledges that goal is "a long way off."

For now, the concrete result is a bounded research demonstration: selected Signal protocol guarantees mathematically verified in four weeks, with human-defined requirements, prior Signal Foundation work as a foundation, and the artifacts released publicly. Whether this becomes a scalable service — or remains a high-effort showcase — depends on economics Google has not disclosed.