Skip to content
Rush Commerce
AI & Automation3 min read

AlphaProof Code verifies Signal: humans still wrote the spec

Google DeepMind's AlphaProof Code produced six million lines of Lean to verify Signal's libsignal in four weeks. Humans reviewed every spec. Copy that split.

AlphaProof Code is Google DeepMind's new agent framework for formal verification, and its first target was Signal. DeepMind announced it on November 10. AI agents wrote more than six million lines of Lean proof about the code under the Signal protocol, in four weeks. DeepMind says expert teams usually need years for work of this type. The useful part for a small team is not the proof. It is the division of labor: people wrote the definition of "correct," agents did the work, and a small trusted checker decided if the work passed.

What actually happened

We read the blog post and the public repository. The facts:

  • The target. The proofs cover compiled AArch64 machine code of libsignal, the official Rust implementation of the Signal protocol. That is the compiled binary, not only the source.
  • Two proofs. The repository has two AI-generated security proofs: one in the symbolic model and one in the computational model. DeepMind calls them independent arguments for security.
  • Who did what. Researchers and models turned Signal's security goals into formal specifications, and human experts reviewed each one. Agents on Gemini Carbon wrote the proofs. Lean checked every proof. DeepMind's diagram labels the prover agents "untrusted" and the gatekeeper "trusted."
  • Human review stayed small. The agents audited the result and flagged about one percent of the codebase for a second look by a person.
  • It catches bad code. The team put two real, previously fixed bugs back into an isolated copy. Both times the system produced a machine-checked proof that the version was insecure.

Signal CTO Ehren Kret said the work "gives us even greater confidence in our code's privacy guarantees."

Read the limits too. The blog says the proofs cover the exact code that runs on devices. The repository is more specific: the artifact is a command-line harness with "hooked cryptographic primitives," so the primitives are trusted, not proved. Neither page mentions side channels or reports a new bug in current Signal. The artifacts are open source under Apache 2.0, and the repository states it is not an official Google product. There is no hosted tool to buy.

Why AlphaProof Code matters for your business

You will not formally verify your order system. You can copy the structure today.

Write the check before the agent writes the code. DeepMind's people spent their time on the specification. In your shop, that is the test suite, the acceptance list, the reconciliation query that must return zero rows. If you cannot state "done" in a form a machine can check, the agent cannot hit it.

Treat the agent as untrusted. DeepMind did, and it built the agent. The proof counts because Lean checked it. Your equivalent is CI that the agent cannot edit, and a deploy step that runs only on green.

Spend review time where the machine points. One percent got human eyes. Ask your coding agent to list what it could not verify, and review that list first.

Test the checker. Put a known bug back and confirm that your tests fail. If they pass, your gate is decoration.

Key takeaways

  • Google DeepMind's AlphaProof Code generated more than six million lines of Lean to verify compiled libsignal code in four weeks
  • Human experts reviewed every specification; Gemini Carbon agents wrote the proofs; Lean checked each one
  • Agents flagged about one percent of the codebase for human review
  • The repository shows limits: a CLI harness with hooked cryptographic primitives, and no side-channel claims
  • For your team: define the machine check first, keep the agent out of the gate, and test the gate with a known bug

An agent is only as good as the check it must pass. We build AI-assisted software with the acceptance tests written first and a CI gate the agent cannot change. See how we build software with checks you own, or tell us which workflow you want an agent to run.

Sources: Google DeepMind, google-deepmind/code-verification on GitHub.

  • #alphaproof-code
  • #formal-verification
  • #ai-agents
  • #lean
  • #deepmind
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.