Google DeepMind says AlphaProof Code formally verified Signal protocol code

Google DeepMind says AlphaProof Code formally verified Signal protocol code

Google DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri describe AlphaProof Code, a framework combining human expertise with AI agents, and say it formally verified key security properties of the code underlying the Signal protocol. They say human experts helped translate Signal’s security goals into mathematical specifications, then AI agents generated more than six million lines of Lean proofs. Lean checks the proofs, while human reviewers examined areas the agents flagged for a second look, described as one percent of the codebase. The authors say the work reaches beyond protocol design to the compiled code running on devices, addressing a gap where vulnerabilities can arise between design and implementation. They report completing the verification in four weeks and testing the system on an isolated version of Signal’s code with two previously fixed bugs; in both cases, it produced a machine-checked proof that the buggy version was insecure. The result is a proof of the specified security properties, not a claim that every aspect of Signal or all deployed software has been verified. The authors say scaling the techniques requires continued research and that they are open-sourcing the verification artifacts. Signal CTO Ehren Kret says the work gives greater confidence in the protocol’s privacy guarantees. Google DeepMind presents the approach as groundwork for other

Publication: 9b6c6627-23b7-4143-8b67-642307a89d51

attribution — pass: The reading attributes the work to Google DeepMind and its named authors, Pushmeet Kohli and Swarat Chaudhuri. It attributes the endorsement to Signal CTO Ehren Kret. Claims stay framed as the source's own statements. Kret's quote refers to 'our code's privacy guarantees', which the draft renders as 'the protocol's privacy guarantees'. That is a close enough paraphrase, but quoting 'our code's' would be more exact.

gives us even greater confidence in our code's privacy guarantees.

summary — uncertain: The final summary sentence is cut off mid-clause ('groundwork for other') and needs completing, e.g. 'for other vital systems, from medical device firmware to power grid software'. The central term formal verification is never put in plain language, though the source supplies one: mathematically proving code runs correctly. Lean and the Signal protocol (which protects private conversations in apps like Signal and Google Messages) are also unexplained. The draft drops the source's 'first time' claims and its 'grand security challenge' framing, which weakens the author's force. Smallest repair: complete the sentence and add the source's own definitions.

Formal verification—mathematically proving code runs correctly without flaws—is the highest standard of security assurance

date — pass: The page dates the post November 10, 2026. This matches evidence.originalPublished (2026-11-10). The feed entry appeared minutes later, consistent with first publication.

November 10, 2026 Research

support — pass: Every claim is supported by the source text and framed as an assertion: Lean proofs of more than six million lines, the 1% flagged for human review, compiled-code scope, four weeks, the two reintroduced bugs, open-sourcing, and the need for continued research. The limitations match the source's 'key security properties' scope. The question follows from the source's own statement that scaling requires further research. Minor: claim 2 omits that the researchers worked with a mix of models when writing specifications, but this is not a contradiction.

Scaling these techniques requires continued research.

media — pass: The selected media is the verified link to the original Google DeepMind publisher page, the most direct access to the source. The metadata title differs from the page headline, but the URL and text identify the same AlphaProof Code post.