🔐Google DeepMind Formally Verifies Signal's Encryption Using AI Agents in Four Weeks

Signal's crypto just got mathematically proven by AI

Google DeepMind just formally verified Signal's messaging security, and the AI did most of the grunt work. The team used AlphaProof Code, a multi-agent framework, to prove the correctness of libsignal's post-quantum PQXDH handshake and Triple Ratchet. Why should you care? Formal verification has always been the expensive kind of rigor. seL4 took about 20 person-years, and CompCert took six years. If you're shipping crypto or security-sensitive code, this is a preview of a world where proofs stop being a luxury item. The catch is that the researchers still had to decide what to prove, and that part stayed human. The numbers: the Signal work took four weeks, a team of five people, and quite a few tokens. Gemini Carbon agents generated 6.5 million lines of Lean code, which then went through audit and human review. The proofs cover forward secrecy, authentication, post-compromise security, and handshake secrecy. Worth watching: the result is theoretical. It assumes perfect cryptography and doesn't guarantee computational secrecy of actual bytes. The team says the next frontier is specification, not proof, and they want to extend this to code in general.