Google DeepMind's AlphaProof Code verifies Signal's cryptography

Google DeepMind introduced AlphaProof Code, an agentic framework that coordinates multiple language models to formally verify software. In a project with Signal, researchers worked with human experts to turn the messaging protocol's security goals into mathematical specifications, then agents built on Gemini Carbon generated Lean proofs checked by machine. The proofs covered the compiled code that runs on devices, not just the protocol design, and the effort took about four weeks. The agents flagged roughly one percent of the codebase for human review. As a test, the team reinserted two previously fixed bugs, and the framework produced proofs that the buggy versions were insecure. DeepMind published a technical paper and is open-sourcing its verification artifacts.