AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Google DeepMind just mathematically verified key parts of Signal

Collected Nov 11, 2026

Google DeepMind researchers have formally verified a set of the Signal protocol's core cryptographic guarantees, using agentic AI to produce mathematical proofs rather than conventional testing. The project, described as a proof-of-concept, took four weeks. Signal's protocol underlies encryption in the Signal app, Google Messages and WhatsApp, so the exercise bears on cryptography protecting a large share of global messaging.

The work rests on formal verification, a technique that uses mathematics to establish that software behaves as specified, rather than sampling behavior with tests. Testing can show the presence of bugs but not their absence; a proof, if the specification and proof are both correct, rules out whole classes of defects within the modeled scope. The researchers used a new DeepMind framework called AlphaProof Code to carry out the proof work. They are releasing the specification and proof they developed for the Signal analysis to support other formal verification efforts.

The analysis found no new vulnerabilities in the portions of the Signal protocol it examined. As a check on whether the method discriminates between sound and unsound code, the researchers ran AlphaProof Code against older versions of Signal containing known, long-fixed vulnerabilities. The framework identified them and showed that those versions could not be formally verified. That negative control is the more informative half of the result for anyone judging the tooling, since a verification system that passes everything is worthless.

Scope is the main caveat. The analysis did not formally verify the entire Signal mobile app; it addressed key cryptographic guarantees of the protocol. DeepMind researchers describe the result as one milestone in a longer process. Swarat Chaudhuri, a DeepMind research scientist, said the team is excited but that a long distance remains before this scales to all the world's software. The researchers also note they benefited from a head start: the Signal Foundation had already done formal verification work on smaller portions of the code, which the DeepMind approach extends.

Ehren Kret, Signal's chief technology officer, said in a statement that prior formal verification covered smaller portions of code while DeepMind's approach enables significant expansion, and that rigorously verifying the implementation gives greater confidence in the code's privacy guarantees as AI-driven security threats grow.

Methodologically, the interesting claim is about process rather than raw proving power. Chaudhuri describes advances in a human-AI collaboration model in which agents contribute and self-correct at nearly every stage while humans are consulted and review at key junctures. He framed the goal as avoiding a setup where AI writes the goals and the proofs and effectively grades its own exam, describing a middle ground instead. In practice, that means the specification still comes from people, and the agent automates the laborious translation of properties into formulas and the construction of proofs, which historically made formal verification too slow and expensive for most projects.

The economics remain unsettled. DeepMind conducted the Signal analysis free of charge as a research project, and Google declined to share its cost. Google also declined to comment on whether it plans a commercial formal verification product or service. That leaves open whether AI-automated verification can scale to large codebases and software supply chains at a workable price, a question the wider push toward AI-aided formal verification has not yet answered.

Pushmeet Kohli, chief scientist at Google Cloud and DeepMind's vice president of research, sketched the upside: if correctness of implementations can be verified, organizations that depend on assurances and worry about bugs, such as banks, could state that they can prove particular properties.

Why it matters: Teams shipping security-critical software, especially in messaging and finance, now have a concrete demonstration that agentic AI can shorten formal verification from a months-long specialist effort toward a weeks-long exercise, though only for bounded parts of a system. The released Signal specification and proof give other researchers a starting point, and the negative results on old, vulnerable versions suggest the approach can flag code that fails its guarantees. The open questions are cost, scale and whether a commercial offering appears, so treat this as evidence of direction rather than a drop-in assurance tool.

Read at Ars Technica · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

The four-week project hints that AI could make formal verification far more practical.