DeepMind unveils AlphaProof Code: AI agents wrote over six million lines of Lean proofs in four weeks to formally verify Signal’s actual code
On 10 November Google DeepMind introduced AlphaProof Code, a multi-model agentic framework for formal verification. Its researchers used it to prove key security properties of the actual code underlying the Signal protocol, which also protects conversations in apps like Google Messages. DeepMind says it is the first time AI has formally verified a widely deployed cryptographic protocol at this depth, and the first time the implementation of such a protocol has been verified this way by any method.
Human experts worked with a mix of models to translate Signal’s security goals into formal mathematical specifications, each reviewed by an expert; agents powered by Gemini Carbon then did the proofs, generating over six million lines of Lean, a language whose proofs are checked automatically by machine. The agents also audited the proof and flagged only about 1% of the codebase for a human second look, with Lean checking the rest.
DeepMind says most protocol verification stops at the design level, while this work proves the security properties down to the compiled code running on devices, such as keeping messages protected before and after a device is temporarily compromised. The verification took four weeks, a task that historically takes expert teams years. To check that the system can tell correct code from buggy code, the researchers reintroduced two real, previously fixed bugs into an isolated copy of Signal’s code, and in both cases the system produced a machine-checked proof that the buggy version was insecure.
DeepMind says it is open-sourcing the formal verification artifacts for community use and has published a technical paper; it sees the approach eventually helping secure critical systems such as medical device firmware and power grid software, though scaling it requires more research.
Sources
- Introducing AlphaProof Code: a breakthrough in verifying complex software at depth — Google DeepMind