DeepMind Uses AI to Formally Verify Parts of Signal’s One-to-One Encryption
Regulatory & Policy · 7 sources · Updated 3h ago · First seen Nov 15, 2026

Google DeepMind researchers used agentic AI to produce computer-checked proofs of selected cryptographic security properties in Signal’s libsignal library. The four-week proof of concept reportedly used 80 AI agents and the AlphaProof Code framework. It covered one-to-one session establishment and message-key updates in compiled AArch64 Linux code from libsignal v0.105.4. The results are not a security certification of the entire Signal application: group conversations, underlying cryptographic implementations and several other components were outside scope, and the proofs depend on stated assumptions about cryptographic primitives.
Separately, Signal disclosed an October US grand jury subpoena seeking information associated with four phone numbers. Only one had a registered account, for which Signal supplied an account-creation timestamp and last-connection date. The disclosure illustrates how limited server-side retention complements message encryption, while the verification work provides additional assurance only for the components examined. Users do not need to change settings in response to the findings. For organizations relying on Signal, independent reproduction and broader verification remain outstanding; the proofs should not be interpreted as eliminating endpoint compromise or other risks outside their scope.


