TechWhoKnows
Log inSign up
Latest
AI
Startups
Venture
Tech
Security
Hardware
Biotech & Climate
Markets & Policy
Companies
TopicsNewsletterAbout
TechWhoKnows

Independent tech news from Singapore, with the sources linked.

Sections

  • AI
  • Startups
  • Venture
  • Tech
  • Security
  • Hardware
  • Biotech & Climate
  • Markets & Policy

TechWhoKnows

  • Companies
  • About
  • Newsletter
  • All topics
  • Search
  • Account

© 2026 TechWhoKnows. All rights reserved.

  • Privacy
  • Terms
Home/AI/Models & Research

Google DeepMind AlphaProof Code Formally Verifies Signal Protocol Machine Code at AArch64 Level

DeepMind has used AlphaProof Code to construct end-to-end mathematical proofs verifying the Signal protocol machine code directly at the ARM64 assembly level.

Google DeepMind AlphaProof Code Formally Verifies Signal Protocol Machine Code at AArch64 Level

By Clara Tan, Artificial Intelligence

14 Nov 2026, updated 15 Nov6 min read

Google DeepMind has demonstrated automated formal verification of end-to-end cryptographic implementations using AlphaProof Code, successfully verifying production machine instructions for the Signal messaging protocol down to the AArch64 (64-bit ARM) assembly tier. The milestone bridges neural theorem proving with systems-level software security, removing the compiler trust gap that historically troubled formally verified software.

Formal verification mathematically proves that a program's execution matches its mathematical specification under all possible execution paths. While researchers previously verified cryptographic algorithms written in high-level domain languages like F* or Coq, compiling those models down to binary machine instructions introduced the possibility of compiler-induced flaws, memory vulnerabilities, or timing side-channels.

Bridging interactive theorem proving and binary analysis

AlphaProof Code extends DeepMind's Lean 4 theorem-proving infrastructure by integrating a formal operational semantics engine for ARMv8-A architecture. The model pairs a reinforcement-learning-guided tree search with automated constraint solvers to generate machine-checked proofs that can be independently verified by Lean's deterministic proof checker.

The verification effort targeted libsignal-core cryptographic primitives, focusing on the Double Ratchet key-exchange algorithm, X3DH (Extended Triple Diffie-Hellman), and Curve25519 scalar multiplication. The neural architecture decomposed the implementation into 14,000 distinct proof lemmas, solving invariant conditions regarding register allocation, stack frame isolation, and branch-free constant-time execution.

By reasoning directly over compiled binary disassembly, AlphaProof Code confirmed that compiler optimisations had not replaced constant-time arithmetic routines with conditional branching logic. Such compiler substitutions have historically created microarchitectural timing leaks capable of exposing private decryption keys to local or remote observers.

Eliminating the compiler trust assumption

In traditional high-assurance engineering, verifying a system requires either building a certified compiler from scratch, such as CompCert, or manually auditing compiled assembly code line by line. Certified compilers frequently produce slower binary output because they disable aggressive loop unrolling and vectorisation passes.

The AlphaProof Code methodology allows developers to compile their source code using standard commercial toolchains, including Clang and GCC with maximum optimisation flags enabled (-O3). Once compilation finishes, AlphaProof Code ingests the resulting machine binary and synthesises a formal mathematical proof establishing equivalence between the compiled binary semantics and the abstract protocol specification.

The entire proof certificate generated for the Signal Double Ratchet binary runs to 180 megabytes of Lean 4 proof terms. When fed to an independent, non-neural Lean kernel, the proof completed validation in 42 minutes on standard workstation hardware, confirming zero arithmetic overflows, zero memory corruption vectors, and strict constant-time execution.

Implications for secure communications infrastructure

The Signal Foundation confirmed receipt of the verification proofs and verified object files. Security researchers noted that extending formal guarantees directly to binary objects drastically reduces the attack surface for zero-click mobile exploits, which often target subtle implementation bugs in cryptographic message parsers.

DeepMind indicated that it plans to release the ARMv8-A formal semantics specification and verification harnesses as open-source tooling for the Lean 4 ecosystem. Future development will focus on verifying cryptographic implementations compiled for the x86-64 and RISC-V instruction set architectures, targeting operating system kernels and hypervisor enclaves.

Frequently Asked Questions

What does binary-level formal verification achieve?

Binary-level verification mathematically guarantees that the compiled machine code running on physical hardware matches the protocol's security specification, ensuring no memory bugs or timing leaks were introduced by the compiler.

Why was the Signal protocol chosen for this verification?

Signal's Double Ratchet and X3DH protocols are the global standard for end-to-end encrypted messaging, protecting billions of users across Signal, WhatsApp, and Google Messages. Verifying its core primitives provides immediate security value to global communications.

Does AlphaProof Code require developers to trust AI outputs?

No. AlphaProof Code outputs formal mathematical proofs written in Lean 4. These proofs are checked by a tiny, deterministic, rule-based verification kernel that rejects any invalid deduction step, ensuring absolute mathematical certainty.

Sources

  1. 1DeepMind Publicationsdeepmind.google
  2. 2Signal Foundationsignal.org

More on

  • DeepMind
  • AlphaProof
  • Cryptography
  • Formal Verification
  • Signal

Further reading

Models & Research

Google DeepMind Unveils AlphaGenome 2 with Single-Cell Resolution for Genetic Medicine

14 Nov 2026

Robotics & Defence

Google DeepMind Unveils Gemini Robotics 2.5 Vision-Language-Action Architecture for Dexterous Manipulation

14 Nov 2026

Models & Research

Google Releases Gemini 4 Hydrogen With 1M Context and Antigravity Orchestrator

13 Nov 2026

Models & Research

OpenAI Dissolves Standalone Robotics Unit to Fold Hardware Work Into Core Models

15 Nov 2026

Follow TechWhoKnows for real-time coverage

Stay updated with the latest developments in technology.

XBlueskyThreadsInstagramLinkedInTikTokYouTube

The TWK Newsletter

Pick your topics and we'll email you our stories on them, with links to the sources.

In this story

Related stories

Google DeepMind Unveils AlphaGenome 2 with Single-Cell Resolution for Genetic Medicine

14 Nov 2026

Google DeepMind Unveils Gemini Robotics 2.5 Vision-Language-Action Architecture for Dexterous Manipulation

14 Nov 2026

Google Releases Gemini 4 Hydrogen With 1M Context and Antigravity Orchestrator

13 Nov 2026