🔍 Intelligence Synthesis JP/EN
Ethics · Regulation · Technology — Pharma Practice Notes

AI Tech Flash ── 2026-11-12

DeepMind Verifies Signal's Core Cryptographic Code

DeepMind says its AI verified Signal's core code in four weeks

What changes

According to Google DeepMind, AI agents mathematically proved security properties of code that actually runs. Earlier verification work mostly examined protocol designs. This work examined the implementation itself.

The company says a task that historically takes expert teams years took about four weeks. That could make mathematical assurance of software security more practical.

Why it matters

Cryptographic protocols underpin nearly all digital communication, such as online banking and web browsing. DeepMind says a single flaw can expose millions of people at once.

The company says real-world vulnerabilities often hide in the gap between design and implementation. This work extends the proofs to the implementation. DeepMind claims it is the first time such a widely deployed protocol's implementation has been formally verified, by any method.

What it is

AlphaProof Code is a framework for formal verification. Formal verification means defining the conditions a program must meet in mathematics and proving that the code meets them. DeepMind calls it the highest standard of security assurance.

The framework is agentic and uses multiple large language models (LLMs). Human experts and AI agents share the work.

Who and where

Google DeepMind researchers built it, including Swarat Chaudhuri and Pushmeet Kohli. The target was libsignal, the core library of Signal's cryptographic protocol.

DeepMind says the protocol is used in apps such as Signal and Google Messages. The company is open-sourcing its verification artifacts. It has also published a technical paper.

Why it was built

DeepMind says formal verification is difficult and often labor-intensive. As a result, much verification stops at the design level.

An implementation can contain flaws that the design does not. DeepMind built this approach to close that gap by automating proof work with AI.

How it works

First, researchers worked with several models to turn Signal's security goals into formal specifications. Human experts reviewed every specification.

Next, agents powered by Gemini Carbon wrote the proofs. DeepMind says they produced over six million lines of Lean. Lean is a computer language that mathematically checks that code is correct.

The agents then audited the proof and flagged areas for human review. These covered one percent of the codebase. Lean checks the rest automatically.

Digital Today reports that the team compiled the Rust-based libsignal into AArch64 Linux machine code. The team then analyzed how that machine code behaves.

Key points

  1. DeepMind says it proved security properties of Signal's core cryptographic code in four weeks.
  2. DeepMind says it reintroduced two previously fixed bugs and proved both versions insecure.
  3. Digital Today reports the verification does not guarantee the safety of the entire Signal app.

Conclusion

Google DeepMind says AI performed implementation-level formal verification in a short time. Digital Today reports the scope is a test program using a specific libsignal version. DeepMind lists medical devices and power grids as possible future uses.

Quotes and sources

“Formal verification—mathematically proving code runs correctly without flaws—is the highest standard of security assurance; however, it is difficult and often labor-intensive.”
── Google DeepMind, 2026-11-10
“Because Lean automatically checks the rest, our experts could focus their review where human judgment matters most.”
── Google DeepMind, 2026-11-10
“In both cases, the system produced a machine-checked proof that the buggy version was insecure.”
── Google DeepMind, 2026-11-10
“The verification does not guarantee the safety of the entire Signal app.”

Source articles

Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.