Google AI Checks Signal's Encryption With Math and Finds It Sound

Google DeepMind says a new AI tool has mathematically verified the Signal protocol, the encryption that protects private messages. The work took four weeks. It found no new vulnerabilities in the parts of the protocol it examined. It is a limited test, not a guarantee for the whole Signal app.

What was checked

The Signal protocol is the set of rules that scrambles messages so only the sender and receiver can read them. It also protects communications on Google Messages and WhatsApp. DeepMind used a tool called AlphaProof Code to do what is known as formal verification. That means using math to prove that software works as intended and has no critical vulnerabilities.

How the test worked

Formal verification has long been seen as a gold standard in software security. It has also been slow. People had to describe what the software should do, turn that into formulas, and then build the proof. AI agents can now speed up the last step, which made a four-week project possible.

The researchers also tested the tool on older Signal versions with known flaws that were fixed long ago. It spotted those problems and showed the protocol could not be verified. That suggests the tool can catch real weaknesses.

What it does not prove

The analysis did not verify the entire Signal mobile app. DeepMind's own researchers call it one milestone in a larger process. They also had a head start from earlier verification work by the Signal Foundation, which covered smaller portions of the code.

Signal's chief technology officer, Ehren Kret, said the approach allows a significant expansion of that earlier work and gives even greater confidence in the code's privacy guarantees.

The bigger goal

DeepMind hopes to go much further. Research scientist Swarat Chaudhuri said there is a long distance to go before this can scale to all the software of the world, and that the team wants to get there. Pushmeet Kohli, a Google executive, pointed to banks and other companies that could one day say they can prove their software has certain properties.

The researchers say humans stay involved. AI agents contribute and correct themselves at nearly every step, but people review the work at key points. Chaudhuri said the team does not want AI writing both the goals and the proofs, and grading itself.

What is not known yet

• Whether this kind of checking can scale to large amounts of code is an open question.
• Google declined to share what the Signal analysis cost. It was done free as a research project.
• Google also declined to say whether it plans a commercial product or service.
• Google is releasing the specification and proof so others can use them in their own verification work.