Google DeepMind Unveils AlphaProof Code, Verifying Signal's Cryptographic Core in Four Weeks

Google DeepMind has unveiled AlphaProof Code, an artificial intelligence system designed to mathematically verify that software strictly meets its security specifications without errors.

In a major demonstration of its capabilities, researchers used the technology to verify the encryption protocol of the secure messaging application Signal. AlphaProof Code completed the formal verification of Signal's core cryptographic code in four weeks—a process that normally takes years of human effort.

Verifying Real-World Code at Scale

According to DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri, previous formal verification efforts often stopped at theoretical designs. AlphaProof Code went a step further by verifying the actual cryptographic code running directly on user devices.

The system relies on AI agents built on Google's Gemini Carbon foundation. After human experts defined the formal security requirements, the AI agents helped create the specifications and generated over six million lines of formal proof in the Lean proof language to independently prove the software satisfied those conditions.

During the process, the AI flagged roughly one percent of the codebase for human review. To evaluate its reliability, researchers deliberately planted two bugs in the codebase for testing, both of which AlphaProof Code caught immediately.

Expanding to Critical Infrastructure

The successful verification of Signal's protocol highlights the growing role of AI in securing sensitive digital systems. Signal Chief Technology Officer Ehren Kret noted that this verification approach helps strengthen trust in data privacy as threat landscapes evolve.

Looking ahead, Google DeepMind plans to expand the formal verification methodology beyond messaging applications to critical real-world infrastructure, such as medical devices and power grids, utilizing systems like CodeMender.