◎THE FACTS AT A GLANCE
  • Google DeepMind's 10 November technical report describes approximately 6.5 million lines of Lean proofs for core properties of libsignal, supported by roughly 26,000 lines of trusted specifications and machine semantics. A Lean kernel checks the submitted proofs; that does not automatically validate every assumption in the specification.
  • The evaluated implementation is libsignal v0.105.4, compiled using a standard Rust toolchain to AArch64 Linux machine code in a dedicated harness. The work covers one-to-one sessions, including the PQXDH handshake and Triple Ratchet, not every Signal platform, feature or group-messaging pathway.
  • Researchers report a four-week verification campaign and tested two previously fixed vulnerabilities by reintroducing them into isolated binaries; the formal security results failed for the deliberately buggy versions. Source artifacts are published for external scrutiny, but the full independent reproduction status needs separate tracking.
01

The Signal

The difference between a secure design and the program that actually runs matters. Cryptography can be proven sound in an abstract protocol model while implementation mistakes introduce new paths for attackers. On 10 November 2026, Google DeepMind presented AlphaProof Code, a human-guided, multi-agent framework designed to close that gap for a widely used messaging protocol.

The system helped formalize security requirements and generated about 6.5 million lines of proof code accepted by the Lean theorem prover. The proofs address authentication, forward secrecy, post-compromise security and key confidentiality in specified models of Signal's PQXDH handshake and Triple Ratchet. The researchers say the campaign took four weeks, compared with the years of expert effort historically associated with comparable industrial proof efforts. That comparison is contextual, not a controlled labor-cost trial.

Importantly, the work reaches compiled machine code from a fixed libsignal implementation rather than just an abstract model written for verification. This is an important technical milestone because production software can diverge from its design. It does not mean a mathematical guarantee now covers all apps, devices, future patches or real-world operating conditions in which Signal-derived protocols are deployed.

02

What the evidence changes

AlphaProof Code separates untrusted work from trusted judgment. AI agents draft specifications with researchers, who approve the properties they intend to prove. Autonomous proof workers then generate Lean code, but the proofs count only if the kernel accepts them and a comparator verifies that they establish exactly the approved specification. Humans do not need to inspect every one of the millions of generated proof lines.

The published technical paper makes the limits unusually visible. The trusted base includes the approved definitions, a model of AArch64 instructions and certain external contracts. Cryptographic primitive security is assumed rather than derived from scratch. The implementation model uses a single-threaded, statically linked Linux process; it does not fully model threads, signals, dynamic linking or external files. The proof applies to libsignal v0.105.4 one-to-one sessions rather than group messaging.

Researchers also reintroduced two known, previously fixed bugs into isolated versions and showed that the corresponding security theorems were false for the buggy binaries. This is an effective validity check: a vacuous or excessively weak specification might otherwise accept software that is visibly unsafe. It is still a test of two historical mistakes, not evidence that every class of real-world vulnerability has been ruled out.

03

The Cortist perspective

Cortist distinguishes a machine-checked proof from a universal security promise. Lean's kernel gives strong assurance that a formal conclusion follows from its premises. Whether those premises capture the right threat model, implementation behavior and operational environment remains a human and empirical responsibility. This is exactly why the authors publish their trusted specifications and semantics rather than treating proof size as a safety score.

Reproducibility is unusually promising here because Google DeepMind is releasing the formal artifacts for external checking. A meaningful independent audit would pin the code revision, Rust compiler, machine-code build, Lean version, proof definitions and dependency hashes; it would then reproduce the theorem checks and examine the threat-model assumptions. A successful external run would strengthen confidence in this scoped verification result.

The work can advance defenders without proving that AI agents should be trusted to make unsupervised security decisions. It complements Cortist's coverage of agent vulnerability discovery: automatically generating candidate exploits and formally verifying specified protections are fundamentally different kinds of evidence. Both require independent verification before their findings become broad deployment claims.

04 / WHAT TO WATCH
  • Independent Lean artifact replay with fixed versions, trusted definitions and documented build hashes, including honest reporting of any reproduction failures.
  • Whether future versions of libsignal, other processor targets and group-messaging functions receive equivalent coverage without silently expanding the original claim.
  • External review of the formal threat model, cryptographic assumptions and trusted AArch64 semantics; distinguish proof correctness from specification correctness.
  • Reproducible evaluations of developer effort, resource costs and mutation tests against additional historically meaningful bugs.
CORTIST TAKE

This is a substantial, inspectable formal-verification result—not proof that every Signal implementation is bug-free. AlphaProof Code's most credible advance is scalable checking of clearly stated security properties, with human responsibility concentrated on what must be trusted.

05

Related evidence and open questions

These research records provide context, not independent reproduction of new claims or proof of a historical case outcome.

Primary sources and verification boundaries

The original publications report a developer-run formal-verification study, an operational control design and a small engineering prototype campaign. They are different evidence types. Cortist has not independently reproduced private tests or certified a hardware product.

Read our editorial standards and research methodology.

Google DeepMind: Introducing AlphaProof CodePrimary announcement, 10 November 2026AlphaProof Code: Verifying Signal's Security at the Machine-Code LevelOriginal 57-page technical paper, 10 November 2026Google DeepMind: public formal verification artifactsPublished proof artifacts and instructions