Google DeepMind announced “AlphaProof Code,” a multi-agent AI framework for software verification. The research team successfully used AlphaProof Code to mathematically prove that the “Signal Protocol,” used for encrypted communications such as Google Messages, has adequate security protections.

AlphaProof Code: Verifying Signal’s Security at the Machine-Code Level
https://google-deepmind.github.io/code-verification/

Formal Verification of Signal’s Protocol with Gemini — Google DeepMind
https://deepmind.google/blog/introducing-alphaproof-code/


Formal verification is the use of formal methods, mathematics, and other techniques to prove the correctness—or identify the flaws—of a hardware or software system. AlphaProof Code is an agentic framework that employs multiple large language models (LLMs), developed to enable formal verification at an unprecedented depth.

AlphaProof Code divides the formal verification work as an actual research team would, determines what needs to be proven, and helps AI agents and human experts collaborate on the proof process. This time, DeepMind’s research team used AlphaProof Code to undertake formal verification of the “Signal Protocol,” an end-to-end encrypted communication protocol adopted by various messaging apps.

In the formal verification of the Signal Protocol, multiple AI models were used to translate the protocol’s security objectives into formal mathematical specifications, all of which were reviewed by human experts.Thereafter, an agent equipped with the state-of-the-art AI model “Gemini 4 Carbon,” featuring advanced cybersecurity and long-term operational capabilities, carried out the proof. Using Lean, a programming language for mathematically proving code correctness, the agent generated a proof spanning six million lines.

The AI agent then examined the Lean proofs and identified the areas that required verification by human eyes. These areas amounted to no more than 1% of the whole, resulting in a substantial reduction in time. Formal verification that would ordinarily have required experts several years was reportedly completed in just four weeks.

The research team further reevaluated AlphaProof Code’s ability to distinguish correct code from code containing bugs by reintroducing two previously fixed bugs into the code. As a result, AlphaProof Code generated verification results showing that the buggy version was unsafe.


Cryptographic protocols such as the Signal protocol underpin nearly all digital communications, from online banking to web browsing. Because a single vulnerability can put an enormous amount of personal information at risk, experts have devoted years to formally verifying these protocols.

However, most protocol verification ends at the design stage, and relatively little progress has been made in verifying the code that is actually compiled. The significance of this effort using AlphaProof Code lies in verifying, down to the precise code level executed on devices, that the Signal protocol satisfies important security requirements.

Signal CTO Eren Kret commented: “Conventional formal verification was necessarily limited in scope, but this new approach makes substantial scaling possible. Previously, verifying that production code precisely matched the cryptographic design required an enormous amount of manual work. Therefore, automatically converting live code into valid Lean theorems using AI agents is a major breakthrough. As AI-powered security threats continue to grow, rigorously verifying the core Signal protocol through automated theorem proving further increases confidence that the code faithfully fulfills its privacy guarantees.”

Pushmeet Kohli, Vice President at DeepMind, said: “Beyond being a breakthrough in research, this is a quintessential example of how cutting-edge AI can serve as an amplifier to counter the pace and scale of modern cyber threats.”