AI Just Proved Signal's Encryption With Math. Here's What That Actually Means

On November 10, Google DeepMind published a mathematical proof that the code behind Signal's encryption actually functions as advertised. This is no ordinary code review or audit; it's a proof, checked by a machine, covering the compiled code that runs on real phones. The project, dubbed AlphaProof Code, involved a team of five researchers and AI agents working under human supervision for four weeks.

Together, they generated over six million lines of Lean proof, which covered libsignal—the reference implementation of the Signal protocol. This protocol safeguards chats in Signal, WhatsApp, and Google Messages, and the work they accomplished would typically take expert teams years to complete.

Most security assurance in the industry stops at the design level. Cryptographers draft a protocol on paper, prove its soundness on paper, and then developers code it. However, this gap often harbors real vulnerabilities. For instance, a ratchet that is flawless in theory could still leak a key if the implementation mishandles certain edge cases.

DeepMind's team didn't verify the paper version; instead, they verified the compiled code, down to the exact instructions running on ARM devices. They proved four crucial properties of that code: message secrecy, correct peer identification, preservation of past messages' secrecy even if a later key leaks, and the resilience of the conversation after a compromise. This last property, post-compromise security, is often overlooked but critical for users.

Significantly, this marks the first time DeepMind has formally verified the implementation of a widely deployed protocol of this magnitude. To validate their findings, they reintroduced two previously identified bugs into an isolated copy of the code. Remarkably, AlphaProof Code produced machine-checked proofs that these buggy versions were indeed insecure. This underscores the effectiveness of their approach.

However, what this proof does not cover is group messaging, which is listed as future work. Additionally, it applies solely to libsignal version 0.105.4, compiled for ARM Linux inside a test harness, not the full Signal app on your phone. The app, servers, key servers, and the user interface—including the safety number display—are all outside the scope of this verification.

Furthermore, the proof assumes the cryptographic building blocks are secure, which includes the low-level cryptographic components like ML-KEM and the curve code. If these underlying math foundations had flaws, the proof would not detect them either.

The implications for developers building messengers are profound. With this breakthrough, the economics of formal verification change dramatically. What once required years for a few experts can now be achieved by a smaller team within a matter of weeks. Open-source artifacts and a publicly available technical paper make this verification accessible beyond Signal.

Additionally, as AI-generated code becomes more prevalent, AI-driven attacks will also accelerate. A defense that once took years to build now takes only four weeks to verify, ensuring a faster and more resilient security posture.