Introducing AlphaProof Code: a breakthrough in verifying complex software at depth

Google DeepMind researchers have developed AlphaProof Code, a new agentic framework that combines human expertise with AI agents to formally verify software at unprecedented depth. The system was used to mathematically prove key security properties of the actual code underlying Signal’s cryptographic protocol, which protects billions of private conversations. This marks the first time AI has been used to formally verify a widely deployed cryptographic protocol at this level of detail, addressing a long-standing challenge in software security.

AlphaProof Code splits verification tasks among AI agents and human experts, translating Signal’s security goals into formal mathematical specifications. The agents then generated over six million lines of Lean code to prove the protocol’s correctness. The verification process took four weeks, a task that traditionally requires expert teams years to complete, demonstrating the system’s efficiency in bridging the gap between protocol design and implementation.

To validate the system’s accuracy, researchers reintroduced two previously fixed bugs into an isolated version of Signal’s code. AlphaProof Code correctly identified both bugs, producing machine-checked proofs that the insecure versions failed to meet security specifications. This confirms the framework’s ability to distinguish correct code from flawed implementations, enhancing trust in its verification capabilities.

The breakthrough enables significant expansion beyond prior formal verification efforts, which were limited to smaller code portions. Google DeepMind plans to open-source the verification artifacts for community use, aiming to improve security across critical systems such as medical devices and power grids. The research was conducted by a team including Swarat Chaudhuri, Andrew Ferraiuolo, and Pushmeet Kohli, with contributions from Signal’s team and other collaborators.