Google DeepMind Used AI to Mathematically Prove Signal's Encryption Works

Researchers at Google DeepMind have formally verified a portion of Signal's encryption code, and they say it is the first time AI has produced a mathematical proof of a widely deployed cryptographic protocol at this depth. The team used AlphaProof Code, an AI system built for formal verification, to prove that parts of the encrypted messaging app's underlying code do exactly what it claims, without errors.

The claim carries weight beyond Signal itself. WhatsApp and Google Messages run the same encryption protocol, so the proof covers cryptography protecting the communications of a large share of the world's messaging users. It also arrives at a moment when AI systems are surfacing software vulnerabilities at a rapid pace — the same class of models that find exploits are now being turned toward proving code correct.

What did DeepMind actually prove?

Formal verification is the mathematical gold standard for software assurance. Instead of testing code against expected inputs, researchers write a machine-checkable proof that the code behaves correctly under all conditions. The catch has always been cost. Developing such proofs manually is slow, labor-intensive work, and the barrier has kept formal verification out of reach for most production software — a "distant promised land," as the researchers describe it.

DeepMind's team, building on earlier verification work by the Signal Foundation, split the task between humans and machines. Humans working with AI agents first identified what needed to be proven. Multiple agents running Gemini Carbon, Google's latest model, then executed the proofs. The agents audited their own output and "highlighted areas that needed a second look by a human — just 1% of the codebase, a massive time savings," the researchers wrote.

The speed gain is the headline number. The entire verification took DeepMind one month to complete what would normally take experts years of manual proof development.

The researchers also stress-tested their system's judgment. They showed AlphaProof Code two previously fixed Signal bugs and asked it to distinguish correct code from flawed code. It produced correct proofs in both cases, demonstrating that the system can tell the difference between sound and broken implementations rather than simply proving whatever it is handed.

Why does this verification matter?

Two firsts anchor the result. "Our work marks the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth," the researchers wrote. "It's also the first time the implementation of a widely deployed protocol like this has been formally verified by any method, with or without AI."

That second claim is the broader one. Even before AI entered the picture, no one had formally verified the implementation of a protocol deployed at Signal's scale by any method. The mathematical argument behind Signal's design had been studied; the shipped code had not been proven correct end to end.

The verification is partial. DeepMind did not verify all of Signal's code. But the proven portion includes core protocol components, and the researchers frame the result as a confidence upgrade rather than a formality. "Having the core Signal protocol rigorously verified with automated theorem proving gives us even greater confidence that our code delivers on its privacy guarantees," they wrote.

For users, the practical meaning is straightforward: the cryptography that Signal, WhatsApp and Google Messages rely on now has a machine-checked mathematical proof, produced in weeks rather than years, that at least parts of its implementation match its privacy claims.

How does this fit the AI-and-security picture?

The result lands in a year — 2026 — marked by an unusually high number of AI-related security incidents, many of them involving agents. AI has spent the period functioning as both threat and defense: models find vulnerabilities in software faster than attackers can exploit them, and vendors are racing to build security-specific models in response. Anthropic, Google and OpenAI are all developing models geared toward cybersecurity, including Anthropic's Claude Mythos.

DeepMind's work extends a broader pattern. AI companies, OpenAI among them, have spent recent years on a run of mathematical breakthroughs, in some cases solving math problems that had stood open for nearly a century. The same large-scale computation that lets models accelerate mathematicians' manual work turns out to apply directly to proof-writing — and proof-writing is exactly what formal verification has always needed.

The DeepMind team itself notes the stakes. The researchers pointed to AI's disruption of the entire cybersecurity space as the backdrop, and positioned this test as a step toward improved security infrastructure rather than an isolated demonstration.

What comes next?

DeepMind has no consumer product planned around the findings, a spokesperson-level caveat the researchers made explicit. But the team is open-sourcing its code so others can apply the same techniques to their own formal verification efforts — a deliberate move to spread the method beyond one company's research lab.

The researchers also named their target list. "Verifying Signal's core cryptography lays the groundwork to secure other vital systems, from medical device firmware to power grid software," DeepMind said.

That is the forward-looking bet: if AI can compress formal verification from a multi-year expert project into a month-long collaboration between humans and agents, the technique becomes practical for the kinds of critical software — medical devices, power grids — where a single implementation error carries physical consequences, and where formal proofs have historically been too expensive to attempt.

(Disclosure: Ziff Davis, which published the original report, filed an April 2025 lawsuit against OpenAI alleging copyright infringement in training and operating its AI systems.)