DeepMind Uses AI to Formally Verify Key Security Properties of Signal’s Protocol Code
Emerging Technology Risk · 14 sources · Updated 1d ago · First seen Nov 10, 2026

Google DeepMind researchers used AlphaProof Code, a human-guided, multi-agent AI framework, to formally verify specified security properties of Signal’s cryptographic implementation. The five-person, four-week project examined the post-quantum PQXDH handshake and Triple Ratchet at the compiled machine-code level, linking protocol guarantees to the libsignal implementation. The properties covered authentication, forward secrecy, post-compromise security, and handshake secrecy. Human experts reviewed the mathematical specifications, while Gemini Carbon-powered agents generated approximately 6.5 million lines of Lean proofs checked by the Lean kernel and flagged roughly 1% of the codebase for additional human review.
Researchers found no new vulnerabilities in the portions examined. In isolated tests, reintroducing two previously fixed bugs caused verification to fail, and the framework produced machine-checked proofs of insecurity. DeepMind is releasing the specification and proof for further research, providing additional assurance for cryptographic technology used by Signal, WhatsApp, and Google Messages. The result is not a security guarantee for the entire Signal application: the verification assumes perfect cryptography and does not establish the computational secrecy of actual message bytes. Defining correct security specifications and scaling verification to larger software systems remain unresolved challenges; security leaders should treat the work as scoped assurance rather than a replacement for implementation testing and broader application security review.


