Google DeepMind's AI proves key Signal protocol guarantees in four weeks

If your product uses encryption like Signal, a proof checked by a computer is a stronger claim than a security audit report. But read what the researchers actually checked before using this claim in a presentation to investors.

Google DeepMind used an AI system called AlphaProof Code to mathematically prove some security guarantees of the Signal protocol. This is the encryption system used by Signal and, with some changes, by Google Messages and WhatsApp. WIRED reported the results this week. The researchers said the work took four weeks and found no new security problems in the parts they tested.

This was a test of an idea, not proof that an entire messaging app is completely safe. According to reports about the WIRED article, DeepMind did not check the whole Signal mobile app. It tested specific promises about how the protocol works, not every part of the apps or every possible way they could fail.

A formal proof is different from a normal code review. In a code review, a person reads the design and decides whether it looks safe. A formal proof is checked by software, step by step, against a written description of what the system should do. The proof is only as reliable as that description and the assumptions behind it. The reports we found do not explain exactly which parts were checked or which software was used to check the proof. We could not confirm DeepMind's own report. So, when someone says "Signal is proven secure," a more accurate meaning is that "these specific security promises hold under these specific assumptions."

One detail makes the result more convincing than a simple success story. Reports about the WIRED article say the team also tested older versions of Signal that had security problems that were already known and fixed. The tool found those problems. It also showed that those older versions could not pass formal verification. This is important because a checking tool should reject code that has known problems.

People were still involved in the process. DeepMind research scientist Swarat Chaudhuri 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."

Why these old security promises still needed checking

Most people did not expect the Signal design to have major problems. Academic studies, including research from Oxford in 2017 on the double ratchet, found no major weaknesses in the protocol. But there has been little proof that the actual code people use follows the design written on paper. That is the problem this work aims to solve.

The Beneficial AI Foundation, which was co-founded by Max Tegmark, is working on this problem through a project called Signal Shot. It uses Lean, a system for checking mathematical proofs, to formally verify the Signal protocol and its Rust implementation. Signal and the Lean FRO are also involved. DeepMind's result tackles the same problem using a different approach.

The most interesting part of this story may be the short timeline. Four weeks is not long for this kind of work. But a proof helped by AI is only useful because separate checking software can test it and reject it if it is wrong.

What this means if you build security software

If you run a crypto wallet, a messaging startup or another business that sells encryption, the main lesson is not that Signal is safe. Many people already believed that. The important point is that computer-checked proofs are becoming affordable enough for people to request them. Customers and security auditors may soon start asking for them. A company that can show a public description of its security promises and a verified proof could have a real advantage over one that only says its software was checked by a respected security firm.

There are also important limits. The Signal protocol is small, has been studied by many experts and is described in public documents. This makes it an unusually good system to test. Most business software is larger and more complicated, and often has no clear written description that can be used to prove it works correctly. Chaudhuri himself says that applying this approach to all software is still a long way off. It is not yet clear whether the cost and effort will be worthwhile for ordinary commercial software.