What is Lean? The proof assistant behind DeepMind's Signal verification, explained

When Google DeepMind announced on November 10, 2026 that its AI agents had verified key parts of Signal's encryption code, the word that did the heavy lifting was Lean. It appears in the first paragraph of the announcement and nowhere in most headlines. If you want to judge whether AI-written proofs deserve trust, you need to know what Lean is and what it can and cannot tell you.

What Lean is

Lean's own website describes it as an open-source programming language and proof assistant that enables correct, maintainable, and formally verified code. Two ideas sit inside that sentence. It is a programming language, so you can write ordinary programs in it. And it is a proof assistant, so you can state a mathematical claim about a program, or about numbers, and write a proof that the software checks step by step.

The front page of the site shows the style. It displays short theorems about number theory and bit operations, and it lists automation tools such as grind and bv_decide, which fill in routine steps of a proof. The site's navigation also points to Mathlib, a library of formalized mathematics, and to a section on program verification. The point of showing both is that the same tool serves mathematicians and software engineers.

Why a proof assistant is different from testing

Testing checks the cases you thought of. A formal proof says something about all cases the statement covers. If a property is "this function never returns a negative number," a test can try a thousand inputs, while a proof covers every input the type allows. That is why formal verification has long been called a holy grail in software security, as WIRED put it, because it uses math to validate that software functions as intended.

The cost has been human time. Before AI, people had to describe what they wanted proved, turn that description into formulas, and then develop the proof. WIRED reported that agentic AI can significantly cut down the last step, which is where DeepMind's AlphaProof Code comes in.

How AlphaProof Code uses it

DeepMind's post says humans decide what to prove and a group of AI agents does the proofs, with the proofs written in Lean. Researchers used a mix of models to turn Signal's security goals into formal specifications, and human experts reviewed every specification. The agents then generated proofs that the compiled code meets those specifications. DeepMind reports over six million lines of Lean, about 1% of the codebase flagged for human review, and a four-week timeline.

The reason Lean matters to the trust story is that the checker is separate from the AI. The agents can be wrong, fast and confident, and it does not matter, because Lean will reject a bad proof. When DeepMind reintroduced two known Signal bugs into an isolated copy of the code, the system produced proofs that the buggy versions were insecure. That works because the checker only accepts what is actually true of the code as specified.

What a Lean proof does not guarantee

A proof is only as good as its statement. If the specification leaves out a property you care about, or states it too weakly, a flawless proof proves the wrong thing. That is why the human review of specifications is the real safeguard. Swarat Chaudhuri, a DeepMind research scientist, told WIRED the team does not want a world where "AI is generating the goals and AI is generating the proofs and basically taking tests that are self-written," and said there is "a middle ground."

Three other limits apply. Scope: WIRED reported the Signal analysis did not verify the entire mobile app and found no new vulnerabilities in the portions covered. Environment: proofs about code say nothing about the operating system, the hardware or the people running it. And cost: Google declined to share what the analysis cost.

Where it fits in the wider AI story

Formal verification is one of the few places where a lab can show a capability that outsiders can audit. That is why it keeps coming up in the race narrative. We covered an earlier case in our report on Claude proving properties of 123 libraries, where the verified property was narrow. The Signal work claims richer guarantees, such as keeping messages protected before and after a device is temporarily compromised, but the same question applies: what exactly was proved?

It also bears on safety. If AI agents are going to write and run large amounts of code, a way to prove properties of that code is a candidate safeguard. It is not a full answer to misaligned behavior, because most unwanted behavior is hard to put in a specification, but it covers the part of the problem that can be written down.

How to read the next claim

When a lab says its AI "verified" something, ask four questions. What was the statement? Who wrote and reviewed it? What proof assistant checked it, and are the files public? And what sits outside the proof? DeepMind says it is releasing the specification and proof for Signal, which lets outsiders ask exactly those questions.

Our position: Lean is the reason this kind of AI claim is more credible than most. A benchmark score depends on the grader's honesty, while a Lean proof depends on a small, open checker. But the credibility attaches to the proof, not to the headline. The statement, the scope and the cost are still claims from the lab until someone else reruns them.