Google DeepMind says it has used teams of AI agents to do something cryptographers have chased for decades: mathematically prove that the code protecting billions of private messages actually does what it claims.

The company's new framework, AlphaProof Code, verified key security properties of the implementation behind the Signal protocol — the encryption used by Signal and Google Messages — down to the compiled Arm machine code that runs on a phone.

The work, detailed in a 10 November research post, is described as the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth, and the first time the implementation of such a protocol has been formally verified by any method, with or without AI.

How the pipeline runs

AlphaProof Code splits formal verification the way a research team would. Human experts work with a mix of models to translate Signal's security goals into formal mathematical specifications — every one reviewed by a person. Then a group of agents powered by Gemini Carbon writes the proofs in Lean, a language whose checker refuses to accept a proof unless it is actually correct. In total the agents generated millions of lines of Lean.

The framework also audits its own output and flags only the parts that need human eyes — roughly one percent of the codebase, DeepMind says — leaving the rest to Lean's automatic checker.

The depth is the point. Most protocol verification stops at the design level, but real vulnerabilities often hide in the gap between a design and the code that ships. The verified properties include guarantees about how messages stay protected before and after a device is temporarily compromised — the scenario Signal's PQXDH handshake and Triple Ratchet are built around.

Four weeks, and a bug-detection test

DeepMind says the process took just four weeks, where expert teams have historically needed years for narrower scopes. To show the system can tell correct code from buggy code, the researchers reintroduced two real bugs that had previously been fixed into an isolated copy of Signal's code; in both cases AlphaProof Code produced a machine-checked proof that the version was insecure.

Signal engineers gave feedback on the methodology, and DeepMind is open-sourcing its formal verification artifacts so others can build on them.

The company frames the result as groundwork rather than a finished guarantee: it says the same approach could eventually be aimed at medical device firmware or power grid software, and presents it as the defensive counterpart to CodeMender, its AI agent for code security, at a time when attackers are themselves using automation.

The caveat is scope. The proofs cover specified properties and the verified codebase — not every part of Signal's stack and not every conceivable attack. DeepMind states that limit in the announcement itself.