Google DeepMind uses AI to mathematically check parts of Signal encryption

In short: Google DeepMind researchers used an AI system to mathematically prove that key parts of the Signal messaging protocol work as intended.

What happened

Google DeepMind researchers ran a four week research project to “formally verify” parts of the Signal protocol. The Signal protocol is the set of rules that helps keep messages private in apps like Signal, WhatsApp, and Google Messages.

Formal verification means using math to check whether software matches its safety and security promises. You can think of it like checking a lock with a ruler and a checklist, not just trying a few keys and hoping it holds.

DeepMind used a new framework it calls AlphaProof Code, along with “agentic AI,” meaning AI that can take steps in a task and correct itself as it goes, with human review at key points. The researchers say their work did not find new security problems in the parts they tested.

To see if the tool could catch real issues, the team also tested older versions of Signal that had known bugs that were fixed long ago. In those cases, the tool flagged the problems and showed that the older protocol could not pass the mathematical checks.

DeepMind is releasing the written specification and the proof it created, so other researchers can build on it. The researchers also noted they could move faster because the Signal Foundation had already done some earlier verification work.

Why it matters

Encrypted messaging is meant to keep private chats private, even from the company running the app. Independent mathematical checks can add confidence in the specific pieces tested, but this did not verify the entire Signal app. DeepMind says a longer term goal is to scale this kind of math checking to much more software, but it is still unclear how practical and affordable that would be.

Source: Wired