Research#Security scan#Code review

DeepMind introduces AlphaProof Code for Signal cryptography

DeepMind says AlphaProof Code used AI agents and human review to verify Signal cryptographic code, with verification materials planned for release.

A blue and purple illustration surrounds a human silhouette with security panels bearing shield symbols

Google DeepMind announced AlphaProof Code on November 10, 2026, describing it as a formal verification framework that combines AI agents with human review. DeepMind says the work checked cryptographic code deployed in Signal and that it plans to open-source verification materials. The announcement does not say the entire Signal app has been proved free of vulnerabilities.

Key points

  • DeepMind says AI agents and human review were used to formally verify cryptographic code deployed in Signal.
  • The company says it plans to release verification materials and points to an accompanying technical paper; we have not verified the paper’s contents here.
  • WIRED stresses that the work concerns part of the Signal protocol implementation, not the whole Signal application.

Background

AlphaProof Code brings a difficult question in software security into focus: how can a team show that cryptographic software meets a security requirement? Formal verification expresses a property as a specification and checks whether an implementation satisfies it. That can produce machine-checkable evidence for a defined claim, but the chosen specification and its boundaries determine what the result covers.

Signal is a widely used encrypted messaging service. Formal verification of its cryptographic implementation ties the claim to properties that can be expressed and checked. According to DeepMind, agents took part in the work and human experts reviewed it. The company says it will release verification materials, which could let others inspect what the proof actually states. That makes the specifications, implementation boundary and proof artifacts more useful to examine than a broad claim that AI found the software secure.

The facts

DeepMind’s announcement says AlphaProof Code combined AI agents and human review to formally verify cryptographic code deployed in Signal. Each part of that description matters. The target is cryptographic code; the method checks specified security properties; and people were involved in the process. The post does not claim that a model independently assessed every security goal, or that Signal’s clients, operating systems or complete product have been proved free of vulnerabilities.

ItemWhat DeepMind’s post says
ProjectAlphaProof Code
TargetCryptographic code deployed in Signal
MethodFormal verification with AI agents and human review
Next stepVerification materials are planned for release; an accompanying technical paper is linked

The planned release matters because verification is easier to evaluate when readers can inspect its artifacts. Security researchers may be able to review the specifications and proofs; software teams can study an example of agents participating in formal methods work. Still, publishing artifacts does not by itself establish that an outside team has independently reproduced the result. We have not treated the linked technical paper as a second, already-checked source: this article has not verified the paper’s full text.

What others say

WIRED describes the work as verification of parts of the Signal protocol implementation and cautions that this is not a proof about the entire Signal app. That distinction is practical. A code path satisfying a given specification says nothing automatically about the app’s interface, device environment, deployment configuration or every possible runtime state. The report adds independent explanation, while details of the technical result should still be checked against the materials DeepMind says it will release.

The Hacker News discussion also focuses on the limits of a proof. Commenters argue that formal verification establishes whether an implementation satisfies the written specification and assumptions; it cannot repair a specification that omits a real security goal or a model that leaves out relevant operating conditions. Those comments are community perspectives, not an independent replication of AlphaProof Code or a substitute for reviewing the proof artifacts. Since DeepMind says those materials will be open-sourced, their clarity and ability to support outside review will matter.

There is a useful comparison with our earlier coverage of AlphaProof Nexus. Both stories involve AI alongside formal proof, but their objects and claims need to be checked against their own sources. For security engineering, a system’s name or the number of agents involved says less than the specification boundary, review process and reproducible artifacts.

Our take

The concrete development is that AI agents have entered a proof workflow for cryptographic code, with human review still part of the process. Teams maintaining security-sensitive software should watch how that division works: agents can help construct and advance formal work, while machine checking tests explicit propositions. That does not turn “some properties were verified” into “the product is completely secure.”

For now, the careful reading is to attribute the result to DeepMind and wait for the materials. Readers should look for the exact code and security properties covered, which specifications were supplied by people, what was excluded, and whether outside researchers can reproduce the checks. WIRED’s scope warning and the Hacker News discussion about specifications both show why “Signal was verified” needs unpacking.

One limit is straightforward: the announcement and a link to a technical paper do not tell us whether a third party can reproduce the work, whether every relevant artifact will be available, or whether the conclusion applies across Signal builds and deployment environments. Formal methods can provide stronger evidence, but only for the properties that were written down and checked. Anyone using AlphaProof Code as a security engineering case study should start with the published specifications and proofs, then decide whether they answer the security questions that matter for their own system.