AlphaProof Mathematically Verifies Cryptographic Foundation in 4 Weeks

Google DeepMind's AI system 'AlphaProof' has mathematically verified the cryptographic foundation code of the messaging app 'Signal'. DeepMind announced completing this process, which typically takes years, in just 4 weeks.

'Mathematical verification' refers to the work of rigorously proving through equations that code is logically correct and free of defects. Unlike bug hunting or functional testing, it mathematically guarantees that code operates as intended under all conditions. While such proofs are especially valued in security-critical domains like cryptography, expert manual work requires enormous time, creating high practical barriers to implementation.

In this effort, a group of AI agents based on Google's large language model 'Gemini' automatically generated over 6 million lines of proof code. Of this, only approximately 1% required human verification, with the remainder handled autonomously by AI. This division of labor, where AI handled the majority of proof work and human experts made final judgments, enabled the short-term completion.

DeepMind has announced plans to apply this methodology to critical infrastructure such as power grids and medical devices. In these domains, system failures directly impact society and human life, making the requirements for software reliability particularly stringent. AI-powered mathematical verification implemented in such domains could substantially reduce the cost and time required to ensure safety.

Signal is a privacy-focused messaging app widely used globally, and its cryptographic protocol is highly regarded by many security researchers and experts. The selection of its core cryptographic foundation as the verification target reflects both its practical importance and technical complexity.

Research into AI-assisted formal verification has been ongoing, but real-world examples applying it to large-scale projects with short-term results remain limited. AlphaProof's automatic generation of 6 million lines of proof code can be positioned as evidence that AI agents are approaching a stage where they can execute specialized tasks requiring advanced logical reasoning beyond simple task assistance.

Future attention will focus on which critical infrastructure domains this methodology is actually deployed to, and how industry and regulators evaluate proofs generated by AI. Mathematical verification, by its nature, has the potential to function as a framework for verifying AI judgments from different angles. It is also reasonable to view AI verification process reliability assurance itself as the next practical challenge.