Google DeepMind said on Tuesday that its AI agents have mathematically proved key security guarantees in the code behind Signal’s encrypted messaging, in a four-week project that once would have taken expert teams years.

The work, announced on 10 November, used a new system called AlphaProof Code to check the real, shipped code of the Signal protocol — not just its design on paper. It is the first time AI has formally verified a widely used encryption protocol at this depth, according to DeepMind, and the first time the working code of such a protocol has been verified by any method at all.

Key Points

  • Google DeepMind announced AlphaProof Code on 10 November, an AI framework that formally verifies software by writing mathematical proofs.
  • Its first test proved key security properties of libsignal, the code library behind the Signal protocol, down to compiled machine code.
  • AI agents powered by Gemini Carbon wrote about 6.5 million lines of machine-checked proof in four weeks.
  • The proof covers the PQXDH handshake and the Triple Ratchet for one-to-one chats — not the whole Signal app.
  • DeepMind found no new flaws in the code it checked, and has open-sourced the proof artifacts for other researchers.

What happened

Formal verification means proving, with mathematics, that software does exactly what it promises — rather than testing it and hoping no flaw was missed. It is the strongest form of security assurance, and also one of the slowest: the work is usually done by hand, line by line, by specialists.

AlphaProof Code splits the job the way a research team would. Human experts decide what must be proved and write the formal specifications, with every one reviewed by a person. Then teams of AI agents, powered by Gemini Carbon, Google’s strongest AI model, write the proofs in Lean, a language built for mathematical argument. Lean’s checker then tests every step and rejects any proof that does not hold.

The target was libsignal version 0.105.4, the open-source library that runs the Signal protocol. DeepMind compiled it with the standard Rust compiler into machine code for 64-bit Arm processors and proved the properties against that compiled code — the exact instructions a device runs — rather than against the design documents above it. That matters because real flaws often hide in the gap between a sound design and the code that implements it.

The proof covers two parts of the protocol: the PQXDH handshake, which sets up a chat and agrees the keys, and the Triple Ratchet, which keeps changing the keys as messages are sent. The properties proved include forward secrecy — past messages stay secret even if today’s keys are stolen — post-compromise security, which locks an attacker out again after a key theft, authentication, and handshake secrecy.

The scale is striking. The agents produced about 6.5 million lines of Lean proof in four weeks, resting on roughly 26,000 lines of trusted specifications and processor rules. Humans did not read those millions of lines: Lean checks each one, so only the small trusted base needs to be right. The agents also audited their own work and flagged about one percent of the code for a second human look.

As a control test, the researchers put two real, previously fixed Signal bugs back into an isolated copy of the code. In both cases the system produced a machine-checked proof that the altered code was insecure. In the code it examined, it found no new vulnerabilities, WIRED reported.

Why it matters

The result is a milestone in two directions at once. For messaging, it is fresh assurance for the encryption that protects billions of private conversations — not only in Signal, but in Google Messages and other apps built on the same protocol family. Signal’s chief technology officer, Ehren Kret, said the work “gives us even greater confidence in our code’s privacy guarantees.”

But the proof has firm limits, and DeepMind and the outlets that covered it were clear about them. It applies to one build of one library, not to the Signal apps on people’s phones. Group chats, key generation, server code and side-channel attacks all sit outside it. The low-level cryptographic building blocks were assumed secure, not proved from scratch. And WhatsApp and other apps run their own versions of the protocol, so the result does not carry over to them automatically.

The second direction is the speed. A landmark earlier success in the field, the formal verification of the seL4 operating-system kernel, took about 20 person-years. AlphaProof Code did comparable depth in four weeks. That changes what formal verification can be used for: DeepMind says the same techniques could eventually secure medical-device firmware, power-grid software and other critical systems, where a single flaw can have physical consequences.

There are open questions, too. Google declined to tell WIRED what the analysis cost or whether it plans to sell verification as a service, and the project was done free as research. DeepMind research scientist Swarat Chaudhuri was cautious about the road ahead, telling WIRED there is “a long distance to go” before the method can scale to all the world’s software. Australia moved this week to set up an AI safety regulator after an OpenAI breach, a reminder that the same technology is being governed as fast as it is being built.

What happens next

DeepMind has published the technical paper and open-sourced the specifications and proofs, so other researchers can inspect them and build on the method. The lab says it wants to apply AlphaProof Code earlier in the design of software, not only after it is written, and to pair it with automated defence tools that find and fix flaws.

For Signal users, nothing changes today: there is no patch to install and no setting to change, because the work verified code that is already shipped. The value is in the assurance — and in the precedent. If AI agents can prove the security of the world’s most sensitive messaging code in weeks rather than years, the same proof could soon be demanded of the software that runs hospitals, grids and banks.

Sources: Google DeepMind, 10 November 2026; WIRED, 11 November 2026; Ars Technica, 11 November 2026; The Register, 10 November 2026; Cyber Insider, 11 November 2026. Published 12 November 2026.