• Sources: primary, discussion, discussion
  • Summary: Google DeepMind introduced AlphaProof Code on 2026-11-10, a multi-LLM agent framework in which humans and models write formal specifications and Gemini Carbon agents generate the Lean proofs, producing millions of lines of proof for Signal's core cryptographic code. Agents audited the proofs and flagged about one percent for human review, and the work took four weeks. When two previously fixed real bugs were reintroduced into an isolated copy, the system produced machine-checked proofs that the buggy versions were insecure, and DeepMind says it will open-source the verification artifacts.
  • Why it matters: Implementation-level formal verification, long too costly outside small cores, now looks feasible for production security code when agents write the proofs and a proof checker validates them.
  • Background: Formal verification proves mathematically that code meets a specification, and Lean is a proof assistant whose kernel mechanically checks every proof step, so a proof an agent writes can be trusted without trusting the agent. Most earlier protocol verification checked the design, not the deployed code, which leaves room for bugs in the gap between them. The Signal protocol underpins end-to-end encryption in Signal and Google Messages, and properties such as protection after a temporary device compromise are among those it aims to guarantee.