Google DeepMind has announced AlphaProof Code, a new variant within its AlphaProof lineage, this time aimed at formal software verification rather than mathematical problem-solving. According to the company, AlphaProof Code agents generated roughly three million lines of proofs written in the Lean proof assistant to verify the core encryption mechanisms used by Signal, the encrypted messaging app.
Formal verification involves mathematically proving that a program behaves exactly as specified, without hidden errors or vulnerabilities, a process that has historically been slow and dependent on scarce human expertise. Automating this task with AI agents capable of producing machine-checkable proofs in Lean would mark a meaningful step for the reliability of critical software, particularly in sensitive areas like cryptography, where a single design flaw can compromise the security of millions of users.
AlphaProof, the system this new version builds on, drew attention in 2024 after reaching a performance level close to a silver medal at the International Mathematical Olympiad, combining a language model with reinforcement learning techniques applied to formal proof search. Extending this approach to source code follows a similar logic: reframing software verification as a proof-search problem that agents can explore at scale.
At this stage, the technical details disclosed remain limited to a brief announcement, without an accompanying academic paper or specifics on the exact methodology, the system's limitations, or how much autonomy the agents actually have during verification. It will be worth watching whether DeepMind releases more comprehensive results, particularly on how reproducible this approach is across other codebases, before fully assessing the significance of this announcement for software security.