SafetyGoogle DeepMind · Signal
DeepMind's AlphaProof Code proved security properties of Signal's protocol code, in six million lines of Lean
6 millionlines of Lean generated to prove Signal's protocol code, in about four weeks
Google DeepMind said on 10 November that AlphaProof Code, a framework of AI agents, proved key security properties of the code behind the Signal protocol in about four weeks. Agents powered by Gemini Carbon wrote over six million lines of Lean, a language a computer uses to check proofs, and flagged about 1% of the work for human review. With two real, fixed bugs put back, it produced proofs the buggy code was insecure. DeepMind calls it a first for a widely deployed protocol. That claim is DeepMind's.