Google DeepMind AlphaProof Code formally verifies Signal protocol implementation

Google DeepMind introduced AlphaProof Code, an agentic framework that formally verified the actual code of the Signal cryptographic protocol, marking the first time a widely deployed protocol implementation has been verified by any method.
The system generated over six million lines of Lean code to prove key security properties, completing the process in four weeks compared to the years typically required for expert teams.
Human experts reviewed only one percent of the codebase, as AI agents handled the bulk of the proof generation and auditing, with Lean automatically checking the remaining work.
The team confirmed the system's ability to detect flaws by reintroducing two previously fixed bugs into an isolated version of Signal's code, resulting in machine-checked proofs of insecurity for both cases.

Technical Methodology

AlphaProof Code operates as a multi-LLM framework where human experts collaborate with AI to define formal mathematical specifications, while agents powered by Gemini Carbon execute the proofs.
The verification process extends beyond design-level analysis to prove that compiled code fulfills security properties, such as message protection during temporary device compromise, down to the exact code running on devices.
The framework uses a workflow where agents draft specifications and proofs, which are then audited by AI to highlight areas requiring human review before final validation by Lean.

Google DeepMind is open-sourcing the formal verification artifacts for community use and has published a technical paper detailing the research.
The project demonstrates how AI can bring elite security assurance to critical infrastructure, with plans to apply these techniques to other systems like medical device firmware and power grid software.