Google DeepMind has hit a software security milestone by using AI to formally verify the security properties of Signal’s cryptographic protocol. Traditionally, cryptographic protocols, the backbone of secure digital communication, are notoriously difficult and time-consuming to analyze formally due to the intricate details involved. This is where DeepMind’s AlphaProof Code steps in, seamlessly combining AI agents with human expertise to tackle the challenge.
AlphaProof is a multi-agent framework that divvies up the verification workload, akin to a team of researchers. Human experts determine the worthiness of specific proofs, while AI agents, powered by Gemini Carbon, handle the actual proving. The result? Over six million lines of Lean code that prove Signal’s implementation is as secure as promised. Impressively, this monumental task was completed in just four weeks, a stark contrast to the years this would typically require from human experts alone.
In a test of accuracy, the framework successfully identified two reintroduced, past bugs, confirming AlphaProof’s capacity to discern secure from insecure code. By addressing the often-overlooked gap between the design and implementation of cryptographic protocols, Google DeepMind’s work not only ensures secure communication apps but also sets a high bar for future formal verification endeavors.
Read more at Google DeepMind…