AlphaProof Code: what it is and what it proved on Signal

AlphaProof Code is a Google DeepMind system where people write down what a piece of software must guarantee and AI agents write the mathematical proof that the code delivers it. In its first public test, announced November 10, it proved security properties of Signal's core encryption code in about four weeks. You can't sign up for it, but the proofs are public.

The post on X

Pushmeet Kohli of DeepMind (@pushmeet) announced it this morning, writing that "we have proven key security properties of the machine code underlying the Signal messaging protocol." He says the work took about a month and that the specifications and proofs are being open-sourced.

Note the words "machine code." DeepMind's announcement says the proofs go "down to the exact code running on devices," meaning the compiled program a processor runs, where a bug in the source or the compiler would still show up.

How it works

DeepMind's announcement describes a split. Human experts turned Signal's security goals into formal specifications and reviewed every one. Agents running on Gemini 4 Carbon then wrote the proofs in Lean, a language whose checker accepts a proof or rejects it. No one has to trust the agents, because Lean checks their work mechanically.

The numbers, from the blog post and the paper:

• Over six million lines of Lean proof (the paper says about 6.5 million), written by agents.
• About 26,000 lines of specifications and processor model that humans have to trust.
• Roughly one percent of the work flagged by the agents for a person to look at.

That ratio is what makes the approach workable: a reviewer reads thousands of lines and leaves the millions to the checker. We covered the same idea in mathematics when Axiom's agents formalized a giant proof in Lean.

The team also tested whether the system can tell good code from bad. They put two real, previously fixed bugs back into an isolated copy of Signal's code. Both times the system proved the buggy version was insecure.

Signal's CTO, Ehren Kret, is quoted in the announcement: "Using AI agents to translate live code into valid Lean theorems is a major breakthrough."

What it proved and what it did not

It would be a mistake to read this as "Signal is verified." The paper is specific about scope.

So the result covers one-to-one messaging through a test harness around Signal's unmodified library. It says nothing about the phone apps, the servers, or your contact list.

How to check it yourself

The announcement doesn't mention a product, an API, a waitlist or a price, so nobody outside DeepMind can run AlphaProof Code yet. The output is public, though, and you can look at it today:

1. Read the paper at google-deepmind.github.io/code-verification. Table 1 lists what each proof covers and leaves out.
2. Browse the Lean code in the code-verification repo on GitHub. The specification files are the short part worth a human's time.
3. If you have a developer on staff, ask them to read the trusted definitions. A proof is only as good as the statement it proves.

DeepMind's own caveat is short: scaling the technique "requires continued research."

Our read

Last week the same lab showed AI agents finding weaknesses in 1024-bit RSA. We read the two together. Agents are getting good at breaking software, and here they are proving a piece of it correct, at a volume of proof no human team would write by hand in a month.

We'd keep expectations modest for now. This is a research result on one carefully chosen codebase. Signal's security goals were already well studied and precisely stated, which made it a good first target, and most business software has never had its requirements written down that carefully. Somebody still has to write the specification, and that part stayed with human experts here.

What interests us is the cost. Machine-checked proof has been too labor-heavy for ordinary software, and DeepMind itself calls it "difficult and often labor-intensive." If agents can do the labor, vendors who handle your customers' data will eventually be asked for more than "we tested it."

What this means for a small business

You can't buy any of this yet, but you can copy how the project divides the work. A person decides exactly what "correct" means, the AI does the volume, and an automatic check at the end accepts or rejects the result. That setup suits invoices and intake forms as much as encryption.

If you want help finding where it fits in your own operation, New Face Design offers a free process audit.