Tendai Gumunyu
Back to journal

AI & Machine Learning

AI Just Started Writing Proofs: What DeepMind's Signal Verification Means for Developers

17 Nov 20268 min readTendai Gumunyu

Google DeepMind's AlphaProof Code used AI agents to formally verify core Signal protocol code in four weeks. What was proven, what was not, and what developers can learn from it.

AI Just Started Writing Proofs: What DeepMind's Signal Verification Means for Developers

Most of us have spent years writing unit tests. On 10 November 2026, Google DeepMind showed what comes after them: AI agents writing mathematical proofs that real code does what it is supposed to do.

The framework is called AlphaProof Code, and its first big target was the cryptographic code behind the Signal protocol, the same protocol that protects conversations in Signal and Google Messages.

Tests versus proofs

A unit test checks that your code gives the right answer for the inputs you thought of. If you test a function with ten cases and all ten pass, you know it works for those ten cases.

A formal proof is different. It shows that a property holds for every possible input, and a program called a proof checker verifies each step. In this project the checker was Lean. If a proof does not hold, it does not pass the checker, however confident the author sounds.

Formal verification has existed for decades, but it has been slow and expensive, usually the work of specialist teams over years. That is the barrier DeepMind is claiming to lower.

What DeepMind actually did

According to DeepMind and the coverage of its paper, the work split into two very different jobs:

  • Humans decide what to prove. Human experts, working with AI models, turned Signal's security goals into formal specifications, and people reviewed those specifications.
  • Agents do the proving. Teams of AI agents wrote the Lean proofs that the compiled code meets those specifications. Reports put the total at more than six million lines of proof.
  • The checker is the referee. Because Lean mechanically verifies every proof, humans did not need to read it all. DeepMind says its experts could focus their review on roughly 1% of the codebase where human judgement matters most.

The verification ran over about four weeks in October 2026, and it worked on the compiled code, not only on the source. To check that the system could tell good code from bad, the researchers put two real, previously fixed bugs back into an isolated copy of Signal's code. AlphaProof Code produced machine-checked proofs that those versions were insecure.

What this does not mean

It is worth being precise here, because "AI proved Signal is secure" is a bigger claim than the evidence supports.

  • It was not the whole app. The work covered key security properties of the core protocol code, not the entire Signal mobile app.
  • The cryptographic primitives are assumed. The low-level crypto routines were represented by idealised contracts, and the paper states that the results are conditional on those being secure.
  • No new bugs were found. Reporting says the analysis did not uncover new vulnerabilities in the parts it examined.
  • It had a head start. Earlier formal verification work on Signal helped, so four weeks is not the cost of starting from scratch.
  • Cost is an open question. Whether this can scale at a workable price has not been answered. Google ran it as a research project and has not shared the cost.

The honest summary: a real step forward in software assurance, not a guarantee of perfect security.

The real shift: the hard part moves to the specification

The most interesting point is where the effort goes. A proof only shows that code matches a specification. If the specification is wrong or incomplete, you have proven the wrong thing very rigorously. So when AI agents take over the proving, the scarce skill becomes deciding precisely what must always be true.

That applies to everyone who ships software, not only cryptographers. As more code is written by AI, the valuable work is defining what correct means and checking it.

What a web developer can take from this today

You do not need Lean to benefit from the mindset. Write down the rules your system must never break, then test those rules, not just examples:

  • Booking systems: a room or tour slot can never be sold twice for the same date.
  • Payments: the amount charged always equals the sum of the order lines, and a confirmed order always has a recorded payment.
  • Multi-tenant SaaS: one customer can never read or change another customer's data.
  • Permissions: a user without admin rights can never reach an admin action, by any route.

Property-based testing tools can then throw thousands of generated inputs at those rules instead of the handful you wrote by hand. It is not a proof, but it moves you in the same direction, and it is especially useful when AI wrote part of the code.

Why it matters now

In the same week, Google also reported attackers using AI to automate the path from a security patch to a working exploit. Faster attacks need stronger assurance on the defending side. Machine-checked proofs, automated bug finding and automated fixing are the defender's answer to machine-speed offence, and DeepMind says it sees this verification work as groundwork for securing other critical systems.

Frequently asked questions

What is AlphaProof Code?
It is a multi-model agent framework from Google DeepMind for formally verifying software. Humans and AI define what to prove, a team of agents writes the proofs, and a proof checker called Lean verifies them.

Does this mean Signal and WhatsApp are now perfectly secure?
No. It means key security properties of the core protocol code were formally verified under stated assumptions. Other parts of the apps and the underlying cryptographic primitives were not covered in the same way.

Will formal verification replace unit tests?
Not any time soon. Proofs are costly and best suited to small, critical cores such as cryptography, payments logic and access control. Tests remain the everyday tool, and the two work well together.

How is a proof different from a test?
A test checks specific examples. A proof establishes that a property holds for all inputs, and a proof checker verifies it step by step.

Building something where correctness really matters?
Get in touch and I will help you work out what your system must never get wrong, and how to test it properly.

alphaproof codeformal verificationsignal protocolgoogle deepmindsoftware testingai agents

Work with me

Have a project in mind after reading this?

Send a short brief and I’ll reply within one business day with a plan, a fixed price and the earliest start date.

Keep reading

All articles