Kivorent News

Wednesday, November 18, 2026

Tech and AINews

Google DeepMind's AlphaProof Code agents formally verify Signal's core encryption code

Google DeepMind said its AlphaProof Code agents, running on Gemini Carbon, wrote machine-checked Lean proofs that the compiled code behind the Signal protocol keeps key security promises. The work took four weeks; DeepMind is releasing the proofs but not the agent system.

Cortex, covering AI models, labs and agents
, 3 min read

4 weekstime DeepMind says the Signal verification took

What to know

  • Google DeepMind said on November 10 that AlphaProof Code, a multi-agent system built on Gemini Carbon, produced Lean proofs about Signal's core cryptographic code down to compiled machine code.
  • Humans wrote and reviewed the specifications of what to prove; agents wrote the proofs, and the agents flagged about 1% of the codebase for a human second look.
  • DeepMind is open-sourcing the specifications and proofs. The agent framework itself is a research project, not a product.

Google DeepMind said on November 10 that a new system of AI agents, AlphaProof Code, has formally verified the core cryptographic code of the Signal protocol, the encryption scheme used by the Signal app and, in variants, by other messaging services. Formal verification means writing a mathematical proof that a program does what it is supposed to do, then having a computer check every step of that proof. DeepMind says the job took four weeks, and that work like this has historically taken expert teams years.

The announcement, by DeepMind's Pushmeet Kohli and Swarat Chaudhuri, calls it the first formal verification of the implementation of such a widely deployed protocol by any method, with or without AI. That is DeepMind's own claim.

How the system works

The split between people and machines is deliberate. Researchers, helped by models, turned Signal's security goals into formal specifications written in Lean, the proof-checking language, and human experts reviewed every one. Agents powered by Gemini Carbon, the model Google opened to paying developers the same day (our earlier story), then wrote the proofs. Other agents audited those proofs and flagged areas for a person to check; DeepMind says that came to about 1% of the codebase. Lean's checker verified the rest.

The proofs apply to compiled code, the instructions that actually run on a phone, rather than only to the source code people write. That matters because a compiler can introduce behavior the source code does not show.

DeepMind's post gives the size of the result two ways: its body says more than six million lines of Lean, while its summary text says more than three million lines of machine-checked proof.

Testing it on known bugs

As a check, the team put two real Signal bugs, both fixed long ago, back into an isolated copy of the code. In both cases the system produced a machine-checked proof that the buggy version was insecure. No new bugs in current Signal code were reported.

Ehren Kret, Signal's chief technology officer, is quoted in DeepMind's post saying the work gives Signal "even greater confidence" in its privacy guarantees.

What it does not prove

An analysis by Tech Insider, drawing on DeepMind's technical paper, says the proofs cover two parts of the protocol, the post-quantum PQXDH handshake and the Triple Ratchet, for four properties: forward secrecy, authentication, post-compromise security and handshake secrecy. Other parts of Signal's library, such as attestation, PIN-based recovery, usernames and media handling, sit outside the proof, as does the rest of the app. The proofs also assume the underlying cryptography is sound rather than proving it. So the result is narrower than "Signal is proven secure," which neither DeepMind nor Signal claims.

DeepMind is publishing the specifications and proofs for others to inspect. It has not released AlphaProof Code itself, and Tech Insider reports it remains a research project rather than a commercial security product.

Why it matters

AI agents that write code have mostly been judged by tests and benchmarks (how to read those), which can miss rare failures. Machine-checked proofs offer a stronger guarantee, and the cost of producing them has been the main obstacle. If agents can do most of that work, as Google also claimed for mathematics with its Antigravity /teamwork results, verification could reach security code that matters to billions of people. The open question is the specifications: proofs are only as useful as the properties humans choose to state. For background on how flaws in widely used code spread, see our software supply chain attack explainer.

Sources

GoogleAI agentsAI models

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.