Google DeepMind's AlphaProof Code uses AI agents to mathematically prove Signal's encryption code is secure
Google DeepMind's AlphaProof Code paired human experts with AI agents to formally verify core Signal cryptography in four weeks. What it means for teams that ship security-critical code.
- Author
Umair Mateen- Published
- Reading time
- 4 min
- Category
- AI
Ideas in this post
Tap one for a 5-second answer. Underlined words in the article work the same way.
0 of 3 understood

Google DeepMind announced AlphaProof Code on 10 November 2026: a framework in which human experts and work together to formally verify software. Its first target was the core cryptographic code of the Signal protocol, the used by the Signal app and Google Messages. We're covering it now because it hasn't appeared on this blog yet and it points to where software assurance is heading.
What formal verification means
Testing checks that code behaves correctly for the cases you try. Formal verification goes further: it is a mathematical proof that the code behaves correctly in every case covered by a written specification. A proof checker then confirms each step mechanically. Until now it has been slow, expert-heavy work, usually done on small pieces of code or on simplified models rather than on the real thing.
What DeepMind says it did
According to the announcement, the work split into stages:
Quick check · Myth or fact
AI agents: myth or fact?
Decide for yourself first, then tap each statement to check it.
0 of 3 checked
From the explainer What is an AI agent? Explained with an assistant you send on errands
- Specifications by people. Humans, helped by AI models, translated Signal's security goals into formal specifications. Human experts reviewed every one.
- Proofs by agents. AI agents built on Google's Gemini Carbon model wrote the proofs in Lean, a language whose checker verifies each proof mechanically. DeepMind describes the result as millions of lines of machine-checked proof.
- Targeted human review. The agents audited their own work and flagged the parts that needed a person to look. DeepMind says that came to about one percent of the codebase; Lean checked the rest automatically.
- Down to what runs on devices. DeepMind says the proofs cover the compiled code, not just a model of it, and that the whole verification took four weeks.
As a sanity check, the team put two real, previously fixed bugs back into an isolated copy of Signal's code. In both cases the system produced a machine-checked proof that the buggy version was insecure. DeepMind calls this the first time a widely deployed cryptographic protocol implementation has been formally verified at this depth, by any method.
What is available
DeepMind says it is open-sourcing the verification artifacts (the specifications and proofs) at google-deepmind.github.io/code-verification, and it links a technical paper, Formal Verification of Signal. The announcement does not say AlphaProof Code itself is available as a product or service. In its own words, “Scaling these techniques requires continued research.”
Who should care
- Teams that ship security-critical code (payments, authentication, encryption, medical or industrial control). Proof-level assurance has been out of reach for most budgets; this suggests AI could make it cheaper, starting with the most sensitive parts of a codebase.
- Anyone relying on Signal or Google Messages. No action is needed. The announcement is about proving properties of existing code, not a fix you must install.
- Buyers and auditors. Expect vendors to start mentioning AI-assisted verification. The useful questions are what exactly was specified, who reviewed the specifications, and whether the proofs cover the code that actually ships.
Limits and caveats
- A proof is only as good as its specification. Here people wrote and reviewed every specification; the AI did the proof work. Anything outside the specification is not covered.
- This was a research project by a large lab on a single, well-studied protocol. Cost, tooling and turnaround for an ordinary company's code are not stated.
- There is no product, price or sign-up today. Treat this as a direction of travel, not something to buy.
What to do this week
- List the few pieces of your software where a bug would be most costly (money movement, login, access control, encryption). That is where formal methods, AI-assisted or not, pay off first.
- If you have engineers interested in this area, have one skim the open-sourced artifacts and the paper to see what a specification looks like in practice.
- Keep relying on tests, code review and security scanning. Verification adds to them; it does not replace them.
- When a vendor claims its code is “formally verified”, ask for the scope and who reviewed the specification.
The DevDrift helps teams add AI to their engineering workflows with human review and built in.
Sources
You’ve read the news. Now understand it.
- Learn · 5 min · Myth or factWhat is an AI agent? Explained with an assistant you send on errandsAn AI agent takes a goal, plans the steps and carries them out with software tools, checking each result as it goes. Explained with an assistant you send on errands.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat is encryption? Explained with a sealed envelopeEncryption scrambles data with a secret key so only someone with the right key can read it. Explained as the difference between a postcard and a sealed envelope.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat is an audit trail? Explained with a visitor logbookAn audit trail is a time-ordered record of who did what in a system, when and with what result. Explained with the visitor logbook at an office reception desk.Start in 5 seconds →


