Proofs about the code, not the design

Google DeepMind said on 10 November that its AI agents had mathematically proved key security properties of the code that implements the Signal protocol, the encryption behind Signal and Google Messages. The company called it the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth, and the first time the implementation of such a protocol has been formally verified by any method, with or without AI.

The distinction matters. Cryptographers have spent decades formally analysing protocols, but DeepMind said most of that work stops at the design level — a proof about the idea rather than about the software people run. In its post, written by Pushmeet Kohli and Swarat Chaudhuri, the company said its agents went further, proving that the compiled code delivers the security properties its experts specified, down to what executes on devices.

Desks and shelves in a quiet library reading room
Human experts wrote and reviewed every formal specification the agents proved against. Illustrative photo. Pixabay · pexels · Pexels License

Among those properties, DeepMind listed keeping messages protected before and after a device is temporarily compromised. Real-world vulnerabilities, it said, often hide in the gap between a design and its implementation, which is exactly what makes closing that gap hard.

How the work was split

The system behind it is called AlphaProof Code, which DeepMind describes as a multi-LLM agentic framework that divides formal verification the way a research team would. Human experts worked with a mix of models to turn Signal’s security goals into formal mathematical specifications, and reviewed every one of them. Agents powered by Gemini Carbon then did the proofs, generating more than six million lines of Lean, the language in which the proofs are machine-checked.

The agents then audited their own output and flagged the parts that needed human judgement — about one percent of the codebase, according to DeepMind. Lean checks the rest automatically. The company put the elapsed time at four weeks, against expert teams historically taking years.

A test with known bugs

To check that the system can tell correct code from broken code, DeepMind’s researchers reintroduced two real, previously fixed bugs into an isolated copy of Signal’s code. In both cases, it said, the system produced a machine-checked proof that the buggy version was insecure.

Signal’s chief technology officer, Ehren Kret, is quoted in the post saying that using AI agents to translate live code into valid Lean theorems is a major breakthrough, and that having the protocol implementation rigorously verified gives the organisation greater confidence in its privacy guarantees.

A network patch panel with cables running into it
The Signal protocol encrypts messages in Signal and Google Messages. Illustrative photo. Brett Sayles · pexels · Pexels License

What it does not settle

Every figure here is DeepMind’s own, published in a company blog post alongside a technical paper; the claim of a first has not been independently adjudicated. A proof also only covers what the specification says, and specifications are written by people — the post is explicit that human experts defined and reviewed each one.

DeepMind said it is open-sourcing its formal verification artefacts, which is the part outside researchers can check. The company framed the result as groundwork for verifying other systems, naming medical device firmware and power grid software, and paired it with CodeMender, the automated code-repair agent it introduced last year. Whether the four-week figure holds on a codebase nobody has spent decades studying is the next thing to watch.