DeepMind verifies Signal protocol with an agentic AI

Google DeepMind formally verified part of Signal protocol’s cryptographic guarantees in four weeks. The exercise validates the method, without yet proving it can extend to the entire software.

Google DeepMind used an agentic AI to mathematically prove several key cryptographic guarantees of the Signal protocol, according to research conducted over four weeks. The experiment matters for the security of encrypted communications, but also for the method: it shows that AI can accelerate formal verification without replacing humans. The work does not cover the entire Signal application, however it has already validated part of the protocol and found known vulnerabilities in older versions.

A useful but still partial demonstration

Formal verification consists of using mathematics to test that software properly respects its security properties. In this case, DeepMind researchers examined portions of the Signal protocol, also used in Signal, Google Messages, and WhatsApp, to confirm essential cryptographic guarantees. The result revealed no new flaws in the studied scope, which strengthens confidence in the protocol’s design.

DeepMind specifies, however, that the analysis does not cover the entire Signal mobile application. It is a step, not an end. The researchers also tested older versions of the protocol containing already known vulnerabilities, and the AlphaProof Code tool correctly identified them, showing it can distinguish correct code from broken code without getting too upset about edge cases.

AlphaProof Code mainly changes execution speed

The core of the announcement lies in AlphaProof Code, a new AI framework designed by DeepMind to assist formal verification. Before this type of tool, precisely describing the expected behavior of software, translating it into formulas, then producing a proof took a huge amount of time. AI does not erase this complexity, but it speeds up the heaviest part of the work and can correct some reasoning errors along the way.

According to Swarat Chaudhuri, a researcher at DeepMind, the goal remains to keep human control at key moments: “We don’t want this sort of vibe world where AI is generating the goals and AI is generating the proofs and basically taking tests that are self-written. There’s a middle ground.” This hybrid approach is also what makes the method credible for sensitive systems, where a validation error would cost far more than a bad PowerPoint presentation.

Why Signal serves as a case study

Signal was not chosen at random. The application already benefits from a history of formal verification work, which helped DeepMind move faster. Ehren Kret, Signal’s CTO, believes that “While prior formal verification focused on smaller portions of the code, Google DeepMind’s approach enables significant expansion.” He adds: “As AI-driven security threats grow, rigorously verifying the Signal protocol implementation gives us even greater confidence in our code’s privacy guarantees.”

For critical services, the challenge goes far beyond encrypted messaging. Pushmeet Kohli, chief scientist at Google Cloud and vice president of research at DeepMind, already envisions uses in banking or other sectors that depend on strong assurances about their code: if one can prove the correctness of an implementation, much of the operational doubt is reduced. However, this promise assumes an uncertain scaling up, especially for very large systems.

The real test will be scaling up

DeepMind has not disclosed the cost of the analysis nor said if a commercial product is planned. The company conducted the exercise as research, free of charge, which limits immediate economic interpretation. The decisive point will therefore be less the technical feat than its reproducibility on more code, at an acceptable cost and with timelines compatible with industry.

The researchers say they have published the specification and proof developed to help other formal verification work. This gesture matters, as the discipline often advances by small verified blocks, not by big slogans. For the sector, the announcement mainly confirms a trend: AI is beginning to become a technical control tool, not just a text or code generator.

Key points

  • DeepMind worked on Signal in four weeks.
  • AlphaProof Code found known vulnerabilities.
  • No new flaw was found in the studied scope.
  • Signal says this approach expands verification.
  • Google did not disclose the cost of the analysis.
  • DeepMind published specification and proof for other work.

By the numbers

  • 4 weeks — project duration, according to Google DeepMind.
  • 3 services — Signal, Google Messages, and WhatsApp use the cited protocol.
  • 1 partial scope — the analysis does not cover the entire Signal application.
  • 2 old versions — tested with known vulnerabilities, according to the researchers.

To read