Google DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri published findings detailing AlphaProof Code, a multi-LLM agentic framework that has achieved the formal mathematical verification of the Signal cryptographic protocol. The protocol, which encrypts private messaging for billions of users across Signal and Google Messages, had long resisted end-to-end implementation proofs due to the vast complexity gap between high-level architectural specifications and production code executing on consumer devices.
The verification workflow operated through collaborative division of labor: human domain cryptographers first worked with foundation models to translate Signal's security invariants into formal mathematical specifications. Autonomous agents powered by Gemini Carbon then executed theorem generation, producing more than six million lines of machine-checked proof in the Lean interactive theorem proving language over a four-week span. DeepMind noted that comparable manual efforts by specialist research teams historically demand years of labor.
To streamline oversight, the agentic pipeline audited the generated proofs to isolate ambiguous segments, surfacing only one percent of the codebase for human expert inspection while Lean's kernel independently checked the remaining ninety-nine percent. The proofs verified that compiled code preserves essential forward secrecy and post-compromise security guarantees, ensuring user communications remain protected even if an endpoint device experiences temporary compromise before or after transmission.
To ensure the system could reliably distinguish valid implementations from subtle flaws, the research team reintroduced two real, previously remediated vulnerabilities into an isolated branch of Signal's codebase. In both validation experiments, AlphaProof Code rejected the flawed logic and generated automated, machine-checked counter-proofs certifying that the modified implementations violated protocol guarantees. Signal Chief Technology Officer Ehren Kret confirmed the milestone, stating that using autonomous agents to translate live code into valid Lean theorems provides unprecedented assurance in Signal's core privacy guarantees.
The research initiative was conducted by Swarat Chaudhuri, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger, Karn Seth, Jonathan Katz, Charles Sutton, and Pushmeet Kohli, with technical feedback from Signal engineers Rolfe Schmidt and Michael Rosenberg. DeepMind has open-sourced the verification artifacts, while noting that verifying more dynamic multi-party protocols and heterogeneous foreign-function boundaries remains active research.