Back to news
researchGoogle2026-11-10

Google DeepMind announced AI-agent formal verification of the Signal protocol, generating over 6 million lines of Lean

Google DeepMind used the AlphaProof Code formal verification framework with Gemini Carbon to translate Signal protocol security goals into mathematical theorems, with agents generating more than 6 million lines of Lean and auditing their own proofs, taking about four weeks.

Google DeepMind announced on November 10 that it had used artificial intelligence to complete formal verification of the security properties of the Signal encryption protocol. The Signal protocol has long been regarded as the gold standard for end-to-end encrypted messaging and, besides the Signal application itself, is also used by WhatsApp and Google Messages, and the research result was described as a milestone. Formal verification means rigorously proving or disproving through mathematics whether a cryptographic algorithm or protocol satisfies its security promises, and unlike traditional testing that looks for known vulnerabilities, it aims to show the protocol meets security requirements under all conditions matching the defined model. Google used an agentic formal verification framework called AlphaProof Code, with researchers and AI models translating Signal security goals into mathematical specifications reviewed by human experts, after which agents powered by Gemini Carbon generated more than six million lines of Lean, a language used to prove code properties mathematically, audited their own proofs and flagged about one percent of the codebase for human review while Lean checked the rest. The team proved that the compiled Signal code met security properties including protecting messages before and after a device is temporarily compromised, going beyond checking the protocol design to verify the exact code running on devices. The whole analysis took about four weeks and Google has not published the specific compute required. Google said it would open-source the formal verification artifacts for community use and described combining formal verification with automated defense agents such as CodeMender, and that it cannot yet determine the practical feasibility of applying the technique more broadly.

DeepMindSignal形式化验证AlphaProof CodeLeanGemini Carbon