Osegi
AI·2 min read

Signal encryption tested by DeepMind's AlphaProof Code AI

Google DeepMind pointed an AI at the Signal protocol for four weeks to mathematically prove parts of it. It found no new flaws in the code it checked.

A smartphone showing the Signal messaging app beside a abstract padlock graphic and lines of mathematical proof code.
Image: The Indian Express
SUMODialoediviewergiftdetect

Google DeepMind has put the Signal encryption protocol through an AI audit, and the protocol held up. Researchers used a new framework called AlphaProof Code to formally verify, meaning prove with maths rather than just test, several key security guarantees in Signal's design.

The work was a proof-of-concept that ran for four weeks. The AI agents found no new vulnerabilities in the parts of the protocol they analysed. When the team pointed the tool at older versions of Signal, it did flag security flaws and could not verify those versions. Those problems were already known and patched long ago.

Why does the Signal protocol matter?

Signal's encryption protocol is open source, and it also underpins WhatsApp, Facebook Messenger and Google Messages. So a proof about Signal's code says something about a lot of people's private chats. Signal also turns on end-to-end encryption by default for all messages and calls, which Telegram does not. Since 2023 it has been upgrading the protocol to resist future quantum computers.

Doing this by hand is slow. Researchers have to describe what the software should guarantee, turn that into formulas, then write the proof. DeepMind says agentic AI can shorten that considerably. Signal's chief technology officer, Ehren Kret, said earlier verification covered smaller portions of the code, and that DeepMind's approach allows a much bigger reach.

What are the limits of the Signal verification?

The scope was narrow. It does not cover the whole Signal mobile app, and it built on earlier formal verification work by the Signal Foundation, the non-profit behind the app. Nobody yet knows whether this can be automated at scale, or at a sensible cost.

DeepMind research scientist Swarat Chaudhuri said there is "a long distance to go" before this scales to all the world's software. He expects a human-AI mix, not an AI writing both the goals and the proofs. DeepMind is releasing the specification and proof for other researchers, and the work cost Signal nothing. Whether DeepMind will sell a verification product is unclear.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:google deepmindsignalalphaproof codeencryptionformal verification

More from this day

See all →

More in: AI

See all →