DeepMind’s AlphaProof Code verified Signal’s cryptographic protocol in 4 weeks

Every developer knows the feeling: the tests are green, the code review is complete, the audit found nothing, and you still cannot sleep soundly. Tests check specific scenarios, and a bug may lurk in one nobody thought of. For an ordinary web application, that may be tolerable. For the cryptography protecting millions of private conversations, it is not.

Google DeepMind is proposing another approach with the release of AlphaProof Code, a multi-agent system that automates formal code verification.

Formal verification, without the math

Imagine a sorting function. Tests might check it against a hundred arrays. Formal verification takes a different approach. First, you express the properties mathematically: the output array is sorted and contains the same elements as the input. You then construct a proof that the code satisfies those properties for any array, rather than a hundred examples.

This provides the highest level of security assurance. Yet hardly anyone uses it: the process is expensive and slow, and specialists are few and far between. The proof often ends up longer than the code itself and has to be written by hand.

What the team did with Signal

The team applied AlphaProof Code to the code implementing Signal’s cryptographic protocol. The protocol powers message encryption in Signal and Google Messages. Agents based on Gemini 4 Carbon wrote more than six million lines of Lean, a language and proof-checking environment for formal proofs. The result is a mathematical proof that the code provides its stated security guarantees.

Two numbers matter most here:

• 4 weeks. According to the authors, a task of this scale usually takes teams of experts years.
• ~1%. That is the proportion of places the agents flagged as needing human attention after auditing their own proof. The rest went through without manual intervention.

Google calls this the first time AI has been used to formally verify a real-world program. That is the company’s own claim; there has been no independent verification yet.

Where is the catch?

First, the limits. A formal proof is only as sound as its specification. If the properties are incomplete or specify the wrong things, the code will be proved correct against the wrong objective. So “security has been proved” should be read as “compliance with specific, explicitly defined guarantees has been proved.” It does not mean “impossible to hack.”

Another caveat: such a proof does not cover everything surrounding the code. The operating environment, hardware, integration errors and human factors remain outside its scope. The source also does not disclose the cost of running the system, its computing requirements or whether it will become publicly available. For now, this is a research demonstration, rather than a product you can plug in tomorrow.

The approach also has a major strength. The proof is checked automatically, so there is no need to take either the agents or the authors at their word. If Lean accepts the proof, it is correct. Model hallucinations are not a concern here: an invalid step simply will not pass the checker.

What to do with this

Developers of critical code (cryptography, payments, OS kernels, consensus protocols): keep an eye on the falling cost of formal verification. If other projects confirm that the work can take a month instead of years, the barrier to entry will drop sharply.

Product managers and founders in security: instead of saying “we passed an audit,” you could present a checkable mathematical proof. That is a compelling argument for regulated industries and enterprise customers.

Everyone else: it is already worth getting used to expressing code requirements as verifiable properties. Once agents learn to write the proofs, the bottleneck will be the ability to state clearly what, exactly, must hold true.

The main takeaway is simple. Agents can already write code, and a way is now emerging to prove its correctness at a scale previously out of reach. The authors believe this may be the only way to counter the speed and scale of modern cyberattacks. That is a debatable claim, but the direction is clear.