Timeline

DeepMind uses AI to formally verify security properties of Signal's protocol code

Gemini agents generated over six million lines of Lean in four weeks; DeepMind said about one percent of the codebase was flagged for human review and Lean checked the rest.

  • Ideas & essays
  • Security & misuse
  • Notable

Google DeepMind introduced AlphaProof Code, an agentic framework in which human experts decide what to prove and AI agents carry out the proofs, and said it had used it to prove key security properties of the code behind the Signal protocol. The protocol protects messages in Signal and Google Messages. The proofs are written in Lean, whose kernel checks each step mechanically.

DeepMind described the process. Researchers worked with a mix of models to turn Signal’s security goals into formal specifications, each reviewed by a human. Agents powered by Gemini 4 Carbon then wrote the proofs, generating over six million lines of Lean (The Register reported 6.5 million). An auditing pass flagged about one percent of the codebase for a human second look, and Lean checked the rest. DeepMind said the work took four weeks, against years for expert teams; The Register reported a team of five people. As a test, DeepMind said it reintroduced two previously fixed bugs into an isolated copy of the code, and in both cases the system produced a machine-checked proof that the buggy version was insecure.

DeepMind said the proofs cover the compiled code running on devices, not only the design, and called it the first formal verification of a widely deployed cryptographic protocol’s implementation by any method. That is the company’s claim. The Register said the verified components were libsignal’s post-quantum PQXDH handshake and its Triple Ratchet, and that the result is “theoretical in that it assumes perfect cryptography”. It did not cover the full Signal app. The work responds in part to a challenge the Beneficial AI Foundation issued in April to verify the Signal protocol and app in Lean.

Swarat Chaudhuri of DeepMind told The Register the process is “fundamentally human-in-the-loop” and that the hard part was deciding what to prove, not checking it. He said the next frontier is specification rather than proof. Humans still reviewed every specification, and DeepMind said scaling the approach needs further research.

DeepMind said it is open-sourcing the verification artefacts and that the approach could extend to systems such as medical-device firmware and power-grid software. It came days after DeepMind and Google Quantum AI reported AI-discovered quantum programs checked in Lean, part of a run of results in which Lean verification is the stated basis for trusting AI output.