# DeepMind's AlphaProof Code proved key Signal guarantees in Lean in four weeks. What it proves, and what it doesn't

> Google DeepMind says AI agents wrote millions of lines of machine-checked proof that Signal's core cryptographic code keeps its security promises, down to the compiled machine code. It is a real milestone for verified software, but it isn't the clean bill of health for WhatsApp that some headlines suggest.

Published Nov 10, 2026 by TechMeasured Staff · AI · Canonical: https://techmeasured.com/ai/deepmind-alphaproof-code-verifies-signal-code-in-lean-what-it-proves

Google DeepMind said on November 10 that a new agent system, AlphaProof Code, produced machine-checked proofs that the core cryptographic code of the Signal protocol delivers key security guarantees. The proofs are written in Lean, and DeepMind says they reach all the way down to the compiled machine code that runs on devices. The whole effort took about four weeks.

Why it matters: formal verification, mathematically proving that code does what its specification says, has been reserved for small, high-value systems because it is slow and expensive. DeepMind says comparable efforts have historically taken expert teams years. If AI agents can do the proof-writing while humans set the goals, verified code stops being a luxury. That matters more as AI-driven attacks find bugs faster than human auditors can.

## What DeepMind says it did

According to DeepMind's post, by Pushmeet Kohli and Swarat Chaudhuri:

- **Humans pick the goals, agents write the proofs.** Agents translated security goals into formal specifications. Human experts reviewed every specification, then agents generated the proofs.
- **The target was real code.** The proofs cover Signal's core cryptographic code, not just an abstract model of the protocol. One of the properties DeepMind highlights is that messages stay protected before and after a device is temporarily compromised.
- **The agents audit themselves, partly.** Agents checked the proofs and flagged about 1% of the codebase for human review.
- **Gemini Carbon powered the agents**, in a framework that mixes several models.
- **The artifacts are public.** DeepMind has open-sourced the specifications and proofs and linked a technical paper.

DeepMind calls it the first AI-based formal verification of a widely deployed cryptographic protocol at this depth, and the first verification of such a protocol's implementation by any method. Those are DeepMind's claims, and no outside reviewer has confirmed them yet.

One number to treat carefully: the post's body says the work produced "over six million lines of Lean," while the page's own summary says over three million. DeepMind hasn't explained the difference.

## Did it find bugs?

No new ones. To show the method can fail a broken implementation, the team reintroduced two real, previously fixed Signal bugs into an isolated copy of the code. In both cases the system produced a machine-checked proof that the buggy version was insecure. Startup Fortune, summarizing Wired's reporting, says the work found no new vulnerabilities in the parts it examined.

Signal was involved. Rolfe Schmidt, Michael Rosenberg and CTO Ehren Kret gave feedback on the method, and Kret called the translation of live code into valid Lean theorems "a major breakthrough."

## What it doesn't prove

Several write-ups have stretched the result, so here are the limits:

| Claim you may see | What the sources support |
|---|---|
| "Signal is proven secure" | Selected security properties of Signal's core cryptographic code, under the stated specifications and assumptions |
| "WhatsApp encryption is proven secure" | WhatsApp uses the same protocol, as 9to5Mac notes, but its own app code was not verified, and 9to5Mac's article body says so even though its headline doesn't |
| "The Signal app is verified" | 9to5Mac says the work covers the protocol rather than the Signal app as a whole |
| "This scales to any software" | Chaudhuri told Wired there is "a long distance to go," per Startup Fortune |

A proof is only as good as its specification. Humans still decide what "secure" means here, which is why DeepMind kept experts reviewing every spec. Signal is also close to a best case: a small, well-studied protocol with clean documentation, as Startup Fortune points out. Messier commercial code will be harder.

Cost is the other unknown. Google declined to say what the project cost or whether it will sell the tool, according to Startup Fortune's summary of Wired. 9to5Mac notes the computing demands are high.

## Who should care

- **Security teams and protocol designers:** the published specs and proofs can be inspected now. They are a template for what "verified" can mean for production crypto code.
- **Anyone choosing a messenger:** nothing changes today. Signal's protocol was already well studied, including Oxford's 2017 analysis of the double ratchet. This adds stronger, machine-checked evidence for the parts covered.
- **Watch next:** whether DeepMind runs AlphaProof Code on other widely used libraries, and whether outside groups check the open-sourced proofs. A separate Lean-based Signal effort, "Signal Shot," is being run by the Beneficial AI Foundation with Signal and the Lean FRO, according to Startup Fortune. Comparing the two will be telling.


## Sources

1. [Google DeepMind — Formal verification of Signal's protocol with Gemini (Introducing AlphaProof Code)](https://deepmind.google/blog/introducing-alphaproof-code/)
2. [9to5Mac — Google DeepMind analysis proves Signal and WhatsApp encryption is secure](https://9to5mac.com/2026/11/10/google-deepmind-analysis-proves-signal-and-whatsapp-encryption-is-secure/)
3. [Startup Fortune — Google DeepMind's AI proves key Signal protocol guarantees in four weeks](https://startupfortune.com/google-deepminds-ai-proves-key-signal-protocol-guarantees-in-four-weeks/)
