Tutorials

Google’s new AI tool demonstrates that Signal’s encryption really works, and here is how.

Google DeepMind used AlphaProof Code with Gemini agents to formally verify parts of Signal encryption, flagging only 1% for human review.

Key Takeaways

  • Google DeepMind used AlphaProof Code to formally verify parts of Signal’s encryption, mathematically showing that selected code behaves as intended.
  • The process combined human expertise with Gemini-based AI agents, flagging only about 1% of the codebase for human review and completing the work in about one month instead of years.
  • The result is a major milestone for AI-assisted software assurance, but it is not a full audit: only portions of Signal’s code were verified, giving confidence rather than absolute certainty.
  • DeepMind is open-sourcing the work to help extend formal verification to other critical systems, such as medical device firmware and power grid software.

Meta description: Google DeepMind used AlphaProof Code to formally verify Signal encryption, flagging just 1 percent of code for human review and speeding up security proofs.

Tags: Signal encryption, AlphaProof Code, formal verification, Google DeepMind, cybersecurity

Featured image: abstract friendly illustration style of a shield, a lock, and small AI agents checking a proof, no real people, no logos, no text

What the verification says about Signal

If you care about encrypted messaging or AI-assisted software assurance, this ZDNET report gives a clear update: researchers at Google DeepMind used AlphaProof Code to formally verify portions of Signal's underlying code. Formal verification is the practice of using mathematical proof to show that software behaves as intended. In this case, the researchers say they mathematically proved that at least parts of the encrypted messaging app's code do what they say, without errors.

That is meaningful for more than Signal fans. WhatsApp and Google Messages use the same encryption, and the work comes as AI systems have discovered many vulnerabilities in software lately, while a large number of people rely on Signal's cryptography to protect communications.

ZDNET summarized the result in three parts: a piece of Signal's encryption code was formally verified; this kind of verification has long been slow and difficult; and AlphaProof Code points toward stronger security checks.

How DeepMind used AlphaProof Code

DeepMind said AlphaProof Code is an agentic framework that combines human expertise with AI agents and performs formal verification at a depth the researchers called unprecedented. The researchers called formal verification the strongest available form of security assurance, but said it becomes hard and time-consuming without large-scale automation, including AI systems.

The broader context is the recent math push across AI labs. Companies such as OpenAI have been making rapid mathematical progress, in some cases solving problems that had remained open for close to a century. Large-scale computation lets AI models speed up much of the manual work mathematicians normally do.

ZDNET disclosed that Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.

To verify Signal, DeepMind built on earlier verification work by the Signal Foundation. The team split the task into two parts. Humans worked with AI agents to identify what needed to be proved. Then multiple agents running Gemini Carbon, Google DeepMind's latest model, executed the proof work. The agents also audited the proof and highlighted areas that needed a second look by a human. The researchers said that was just 1% of the codebase, saving a huge amount of time.

The team also tested AlphaProof Code's ability to tell correct code from flawed code. Researchers showed it two previously fixed Signal bugs, and AlphaProof Code produced correct proofs for both.

Why the result is a milestone, but not a finished audit

The researchers said no prior work had used AI to carry out formal verification of a widely deployed cryptographic protocol at this level of depth. They also said no earlier method, AI-based or human-based, had formally verified the implementation of such a widely deployed protocol.

That is a notable threshold because formal verification normally requires humans to manually develop a proof that software meets a specification. That heavy manual effort has made formal verification feel like a distant promise rather than routine practice.

DeepMind's team said the work took one month, where expert manual effort might normally take years. Still, the work did not verify all of Signal's code. DeepMind said the test is a step toward improved security infrastructure at a time when AI is reshaping the entire cybersecurity field.

For Signal, the implication is confidence rather than certainty. The researchers said rigorous verification of the core Signal protocol through automated theorem proving gives them greater confidence that the code meets its privacy promises. Automated theorem proving uses software tools to construct or check mathematical proofs about code behavior.

What this means for cybersecurity

The result also captures a split reality in cybersecurity. AI is both a large risk and a powerful remedy. ZDNET said 2026 has been especially notable for AI-related security incidents, many involving agents. At the same time, Anthropic, Google, and OpenAI are building models focused on cybersecurity, including Anthropic's Claude Mythos.

In that setting, DeepMind's Signal work is not just a product update. It is evidence that AI agents can handle parts of a proof process that previously depended on scarce expert time. It also shows the limits of a single result: partial verification is valuable, but it is not the same as proving an entire messaging system.

DeepMind did not announce plans to turn the findings into a consumer product. Instead, the researchers are open-sourcing their code to help others carry out formal verification. They said verifying Signal's core cryptography creates a template for protecting other essential systems, including medical device firmware and power grid software.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.