Google DeepMind says its AlphaProof Code agent framework formally verified key security properties of Signal's protocol implementation in Lean

Google DeepMind says its AlphaProof Code framework, in which humans review formal specifications and Gemini Carbon-powered agents write the proofs, verified key security properties of the compiled Signal protocol code. The properties include message protection before and after a temporary device compromise. The work reportedly took about four weeks and produced over six million lines of Lean proof, with only about 1% of the codebase flagged for human review. As a sanity check, two previously fixed real bugs were reintroduced, and the system produced machine-checked proofs that the buggy versions were insecure. DeepMind has published a technical paper and says it is open-sourcing the verification artifacts.

If the results hold up, this shows that AI agents can cut the cost and time of formally verifying real, deployed security code, which has usually taken specialist teams years. Cryptography and security engineers, and anyone maintaining high-assurance software, are the people most affected. The practical effect is limited for now. This is a research result on one well-structured protocol, not a tool teams can buy or run on their own code. Because Lean proofs can be machine-checked, outside researchers can test the claims, which makes this more credible than a typical vendor announcement.

The "breakthrough" and "grand challenge" framing and the "first at this depth by any method" claim are DeepMind's own and need independent confirmation. The talk of securing medical devices and power grids is speculation, since nothing here shows it transfers beyond Signal. A proof is only as good as the human-reviewed specification, so "verified secure" means "matches the spec", and the 1% human review is where mistakes could hide. The four-week figure also leaves out the cost, the compute and any earlier verification work the team built on.

1. 01Independent review of the released Lean artifacts and paper: whether outside verification experts confirm the specifications cover the properties that matter, and whether the proofs check cleanly.
2. 02Whether Signal, or another widely deployed project such as Google Messages, adopts the approach in its own build or review process, and whether it finds any new bugs rather than only reintroduced old ones.
3. 03Replication on a different codebase, with published cost, compute and human-hours, to show whether the four-week result generalises or was specific to Signal.