AlphaProof Code

AlphaProof Code is an agentic formal-verification framework announced on 10 November 2026 by Google DeepMind. In a blog post signed by Pushmeet Kohli and Swarat Chaudhuri, the lab says researchers guided AI agents to mathematically prove key security properties of the actual machine code underlying the Signal protocol — the cryptography that also protects chats in apps such as Signal and Google Messages. DeepMind vice president of research Pushmeet Kohli posted the announcement as @pushmeet.

The system is described as a multi-LLM framework that splits formal verification the way a research team would: human experts work with models to decide what to prove, then a group of agents does the proving. DeepMind says researchers first translated Signal's security goals into formal specifications (every specification reviewed by humans). Agents powered by Gemini Carbon then generated the proofs in Lean — more than six million lines, in the lab's account — and flagged about one percent of the codebase for a second human look. Lean machine-checks the rest.

The lab calls this the first time AI has formally verified a widely deployed cryptographic protocol at this depth, and the first time the implementation of such a protocol has been formally verified by any method. In DeepMind's telling, most prior protocol work stops at the design; the agents here aimed at compiled code, including properties such as protection before and after a device is temporarily compromised. The verification is said to have taken four weeks, a task the post says historically takes expert teams years. As a check, the team reintroduced two real, previously fixed bugs into an isolated copy of the code; in both cases the system produced a machine-checked proof that the buggy version was insecure. Signal CTO Ehren Kret is quoted calling the translation of live code into valid Lean theorems a major breakthrough, while noting that earlier formal work had focused on smaller portions of the code.

Highlights:

• Official blog (10 November 2026): https://deepmind.google/blog/introducing-alphaproof-code/ ;
• Official announcement: https://x.com/pushmeet/status/2120127611137634376 ;
• Distinct from the earlier AlphaProof Nexus line of work (Erdős / Lean proofs). DeepMind presents AlphaProof Code as a new framework for verifying software implementations;
• The blog says formal verification artifacts are being open-sourced; it also mentions pairing this line of work with automated defense agents such as CodeMender in future research. Axios, reporting the same day, quotes the team describing the feat as a research project, not a commercial cybersecurity product.

Why it matters

Formal verification — proving that code meets a mathematical specification — is the highest bar commonly cited for software assurance, and it has usually been too expensive to apply at the depth of a widely deployed protocol implementation. AlphaProof Code is DeepMind's bid to use agent teams, with humans still writing and reviewing the spec, to reach that depth on production code. If the artifacts check out, it is a concrete data point that frontier models can generate Lean proofs large enough for a real cryptographic stack, and that a small human audit surface can be enough when a kernel such as Lean checks the rest.

For people who work on secure messengers, protocol implementations, or AI-assisted verification, the useful objects are the official write-up, the specification, and the promised open artifacts — not the headline that Signal is now "proven secure" in every sense. DeepMind is explicit that this is selected properties of the core cryptographic code, not a full audit of the Signal app.

Limits

The claims, line counts, four-week timeline, and bug-reintroduction check come from DeepMind (and contemporaneous reporting). We have not independently checked the Lean artifacts. Coverage is of specified security properties of the protocol implementation, not of the entire application, side channels, operational security, or future code changes. Formal proofs also assume their models of the hardware and cryptography; a gap in the specification is a gap in the guarantee. The announcement does not present AlphaProof Code as a generally available product or API. Treat the name as a research framework until DeepMind ships a supported tool.

1. Read the official post https://deepmind.google/blog/introducing-alphaproof-code/ and the announcement https://x.com/pushmeet/status/2120127611137634376 ;
2. When DeepMind publishes the formal-verification artifacts mentioned in the blog, inspect the specification and Lean proofs rather than relying only on the summary;
3. Keep this distinct from AlphaProof Nexus, which was a different DeepMind effort on mathematical proofs, not this software-verification framework;
4. Do not treat the result as a blanket guarantee for Signal clients, or as a product you can point at an arbitrary codebase today — it is a research milestone on a specified slice of a protocol implementation.