JOURNAL / 2026.11.11

AlphaProof Code verifies cryptographic properties of libsignal: what the proof guarantees

DeepMind connects Signal security properties to machine code using agents and Lean. The advance broadens what can be verified, while shifting work toward deciding exactly what must be proved.

What would an AI have to deliver for us to trust a program's security? AlphaProof Code, introduced by Google DeepMind on November 10, offers a concrete answer: checkable proofs about existing code. The project combines experts and agents; agents powered by Gemini 4 Carbon generate the proofs and Lean checks them. Human involvement includes reviewing the specifications: the statements defining what it means to meet the goal.

The original research connects cryptographic guarantees for libsignal to an AArch64 executable. It addresses PQXDH session establishment and Triple Ratchet. The five-person team worked for four weeks and spent most of its effort on specifications. The scope has important boundaries: it studies a build of a command-line tool, rather than Signal's complete applications; several routines are represented by assumed contracts. Lean checks deductions, but the faithfulness of those premises requires review.

This extends a story we previously covered with Claude and Theorem. That work focused on guarantees concerning execution failures in libraries. Here the question is what an adversary can learn or forge while interacting with a protocol. My reading is that the meaningful change lies in the richness of the property: proving that a program respects a security objective requiring reasoning about messages, keys and successive interactions.

The computational proof guide specifies that it concerns one-to-one sessions and does not incorporate the post-quantum theorems from its reference works. This is an essential caution when reading names such as PQXDH: using a component designed for quantum resistance does not automatically turn every proof about it into a guarantee against quantum adversaries.

There are tools for examining that distance between headline and result. The paper-to-code map links claims to Lean declarations and distinguishes proved, conditional and open results. It also lists restrictions, such as maximum input sizes. For someone maintaining a dependency, that map is more useful than a count of generated lines: it identifies which claim supports a decision and which conditions could invalidate it.

The public repository documents building and checking, with pinned tool versions. It includes a comparator that checks that theorems match the trusted statements, including their transitive definitions, and checks the proofs again in a fresh kernel. This addresses a practical problem: proving something easier after changing a definition can produce a valid proof of the wrong goal. Available artifacts enable scrutiny; they do not mean that we have reproduced the result.

To build with this technology, I would start with a bounded question with visible consequences. For example, in a shared-document service, I would want to specify when revoking access must prevent a read and what happens to sessions already open. Making those decisions explicit could reveal disagreements about the product before any agent attempted a proof. This is our own example, not an application tested by DeepMind.

I would then ask for the guarantee to accompany the program: exact version, statement, assumptions and checking procedure. An update changing session handling would require reviewing which proofs remain relevant. I would budget for that review, too; the publication does not establish the cost of maintaining equivalent guarantees in an ordinary service.

In my view, this offers a more lasting opportunity than automating another stage of programming: making security decisions reviewable when they are currently scattered across documents and code. The risk would be turning “verified” into a label that excuses us from reading the statement. I would adopt the method for the precise questions it forces us to answer and the artifacts it leaves for discussing them.

Primary sources consulted on November 11, 2026. We have not run the verifier or reproduced the proofs. The editorial comparison and adoption proposals are our own analysis.

← Back to journal