Knowledge Engine

DeepMind says AI agents proved security properties of Signal's compiled code: what the proofs cover

·no read data

What happened

On 10 November 2026 Google DeepMind announced AlphaProof Code in a blog post by Pushmeet Kohli and Swarat Chaudhuri, alongside a technical paper with nine Google authors and a public release of the proof files. The target was libsignal, the open-source library that implements the Signal messaging protocol used by Signal and other apps.

According to the paper, the team compiled libsignal version 0.105.4, with a small command-line harnessevaluation harness The software setup that feeds a test to an AI model and scores its answers, whose settings can change the results. around the unmodified library, into a statically linked AArch64 Linux program. Agents then produced proofs in the Lean proof assistantproof assistant A computer program that checks every step of a mathematical proof and accepts it only if each step follows from the rules., about 6.5 million lines in total, that this compiled machine code satisfies four security properties for one-to-one sessions: forward secrecyforward secrecy A guarantee that if someone steals a device's keys today, they still cannot read messages sent before the theft., post-compromise securitypost-compromise security A guarantee that after a device's keys are stolen, later messages become secret again once the protocol refreshes its keys., authentication and handshake secrecy. The verified operations are setting up a session, encrypting and decrypting. Key generation was not verified.

The blog says the proofs were written by agents running Gemini Carbon, and that specifications were drafted with a mix of models and reviewed by human experts. The paper describes campaigns of 40 to 80 agents on a single machine for 24 to 72 hours; the blog says the whole project took four weeks. As a check, the team reintroduced two previously fixed libsignal bugs into an isolated copy, and the system proved that version insecure.

What is claimed vs. what is verified

The central result is unusually checkable. A Lean proof is accepted only if Lean’s small kernel checks every step, so whoever or whatever wrote the proof matters less than for most AI output. DeepMind released the artifacts, which means outside researchers can rerun the checker. As of 12 November, none of the coverage reviewed for this post reports an independent group doing so. A detailed reading of the paper by developer Nikoloz Turazashvili on dev.to says plainly that he did not rebuild the proofs.

Several statements remain DeepMind’s own. The company describes this as the first time the implementation of a widely deployed protocol has been formally verifiedformal verification Using mathematical proof, usually checked by a computer, to show that a program does exactly what a written specification says. by any method. The paper phrases it more narrowly, as the first proofs connecting protocol-level guarantees to unmodified code. The blog says human experts reviewed every specification and that only one percent of the codebase needed a second human look; those review figures have not been independently assessed. Google declined to share the project’s cost, according to The Next Web, citing Wired.

What a reader should hold onto is what the proofs rest on. The paper lists a trusted basetrusted computing base The parts of a system that are assumed to be correct rather than checked, so any error in them could undo the guarantees. of about 26,000 lines: the theorem statements and definitions, plus a Lean model of 467 AArch64 instructions and 31 system calls that the team tested against other processor simulators. The cryptographic primitives themselves are assumed to work. The Register’s Thomas Claburn calls the result theoretical for that reason. The paper also lists what is outside scope, including group messaging, multiple devices, side channels and quantum adversaries in the computational proof, and notes that its computational security bound becomes vacuous for many long sessions.

Why it matters

Formal verification of real systems has historically been slow. The Register reports that the paper compares its four weeks with roughly 20 person-yearsperson-years A count that adds up how many years each person in a study was followed, so that people watched for different lengths of time can be combined. for the seL4 operating system kernel and six years for the CompCert compiler. If agents can write the proofs while humans write and check the specifications, the bottleneck in verifying software moves from proof labour to specification review and compute. Proving properties of compiled machine code, rather than source, also means the compiler does not have to be trusted, since the paper says its proofs cover the compiled artifact rather than the Rust source or the compiler.

For Signal users, the practical meaning is narrower than the headlines. The verified artifact is a Linux build of the library through a test harness, not the phone apps, and The Next Web summarises the guarantees as covering one-to-one chats, not the Signal apps.

Open questions and other views

Signal’s chief technology officer, Ehren Kret, is quoted in DeepMind’s blog calling the work a major breakthrough. That endorsement is not independent: the acknowledgments thank three Signal staff for feedback on the methodology.

Sceptics can point to the specifications. A proof shows that code meets a statement; if the statement misses an attack, the proof does not help. DeepMind’s answer is human review of every specification, though the paper itself says such review is not expected to catch everything. Turazashvili also notes that checking a theorem’s text is not enough when the definitions it relies on can change, a risk the project addresses with a separate comparison tool.

Others will ask about cost and generality. The Signal protocol already had well-studied security goals to write specifications from; it is unclear how the method would do on software whose intended behaviour is less clearly defined. Whether this becomes a product is also open: Google did not say, according to TNW’s account of Wired’s reporting.

Sources

  • Google DeepMind (Kohli, P. & Chaudhuri, S.), “Introducing AlphaProof Code”, 10 November 2026. deepmind.google
  • Chaudhuri, S. et al., “AlphaProof Code: Verifying Signal’s Security at the Machine-Code Level”, technical paper and artifacts, Google DeepMind, 10 November 2026. google-deepmind.github.io
  • Claburn, T., “Google DeepMind melds humans and AI agentsAI agent An AI system that carries out multi-step tasks on its own, such as running code or using tools, rather than only answering one message at a time. to vet Signal’s secure messaging techniques”, The Register, 10 November 2026. theregister.com
  • Constantin, A. M., “Google DeepMind’s AI agents formally verify Signal’s encryption code”, The Next Web, 12 November 2026. thenextweb.com
  • Turazashvili, N., “Google DeepMind’s Signal proof: what AlphaProof Code checks”, DEV Community, 10 November 2026. dev.to

Check yourself

Answers are checked in your browser. Nothing is recorded or sent anywhere.

  1. What exactly did DeepMind's proofs cover, according to the technical paper?

  2. What did the reintroduced-bug experiment show?

  3. Which of these is assumed rather than proved, according to the paper?

  4. A reader says: 'The 6.5 million lines of proof were written by AI, so we have no reason to believe any of it.' Which response fits the evidence best?

Questions without answers

  1. If a proof shows a program meets its specification, but humans wrote the specification, where should a security reviewer now spend their time?
  2. Google has not disclosed what the project cost. What would you want to know about cost before judging whether this approach could be used on ordinary software?

Back to the archive