Atlas AI Resources

← Back to AI News

AI and security

Google DeepMind used AI to prove key parts of Signal’s encryption code are secure. Here is what that covers.

Published November 12, 2026

On November 10, Google DeepMind said a team of its AI agents had written a mathematical proof that core parts of Signal’s encryption code do what they promise, such as keeping old messages secret even if a phone is later hacked. Signal’s encryption protects chats in the Signal app, Google Messages and WhatsApp. The proof was checked by a computer, not taken on trust. It is a real step for software security, but it covers specific pieces of code, not the whole Signal app.

What happened

The work uses a method called formal verification. Instead of testing software with lots of inputs and hoping nothing breaks, you write down exactly what the code must guarantee (that written rule is called a specification) and then prove, step by step, that the code always meets it. WIRED, in a story republished by Ars Technica, notes that fans call it a holy grail of software security. It is also slow and expensive, so it is rare.

DeepMind built a system it calls AlphaProof Code. People and AI first turned Signal’s security goals into specifications, and DeepMind says human experts reviewed every one. Then a group of AI agents, which are AI programs that work through a long task on their own, wrote the proofs. DeepMind says the agents ran on its Gemini Carbon model. The proofs are written in Lean, a programming language built for math proofs. Lean’s checker rejects any proof with an invalid step, no matter who or what wrote it.

The paper says the agents produced about 6.5 million lines of proof. That rests on about 26,000 lines that a reader has to trust, mostly the specifications and a description of how the phone’s processor runs code. The proofs cover Signal’s opening handshake, where two phones agree on a secret key, and its “ratchet,” which makes a fresh key for every message. They show properties like forward secrecy (stealing today’s keys does not unlock older messages) and post-compromise security (once a hacker loses access, new messages become safe again).

The proofs are about machine code, the low-level instructions a processor actually runs, built with the normal compiler from Signal’s unmodified code library. Most past work only checked a protocol’s design on paper. As a test, the researchers put back two real bugs that Signal had already fixed. In both cases the system proved the buggy version was insecure. The paper says a team of five did the work in four weeks.

Why it matters to a normal person

Signal’s protocol code now has a published machine-checked proof for one-to-one chats. It covers the reference version of the code in libsignal, Signal’s code library, not the apps’ own builds. The work found no new flaws in the parts it examined, Ars reports. Signal’s chief technology officer, Ehren Kret, said in a statement that it gives the company “even greater confidence in our code’s privacy guarantees.”

The bigger point is cost. If AI can do the slow proof work, this kind of guarantee could reach software that never had it. DeepMind names medical device firmware and power grid software as future targets.

What it does not mean

It does not mean Signal is proven unhackable. The paper verifies one-to-one chats only and leaves group messaging for future work. The basic encryption building blocks were assumed to work, not proved. Side channels, meaning leaks through things like timing, are out of scope, and so is having Signal on several devices. The researchers checked their own test build of the code, not the Android or iOS apps, and Ars notes the work did not verify the whole mobile app.

It also does not mean AI did this alone. People decided what to prove and reviewed every specification. DeepMind researcher Swarat Chaudhuri told WIRED they do not want a world where “AI is generating the goals and AI is generating the proofs.” Ars adds that DeepMind had a head start from Signal’s own earlier verification work, and Google would not share what the project cost.

Finally, this is DeepMind’s own technical paper, posted on its website rather than in a peer-reviewed journal. The proofs and specifications are public, so other experts can check them.

Sources