Google DeepMind said on 10 November that AI agents in a new framework, AlphaProof Code, proved key security properties of the compiled code behind the Signal protocol. The agents generated about 6.5 million lines of machine-checked Lean proof code, resting on about 26,000 lines of trusted specifications and AArch64 semantics, according to the technical paper. DeepMind’s blog post says the verification process took four weeks.
The paper says the proof work was done by a team of five over four weeks, which is 20 person-weeks. It cites the formally verified seL4 microkernel at twenty person-years, which is 1,040 person-weeks, 52 times as many, by our arithmetic. The two systems differ in size and scope, and the five researchers ran dozens of AI agents alongside them.
Business Pill · Verification Cost
An evergreen Business Engineer explainer on verification cost. It teaches the general idea only and does not discuss DeepMind’s paper.
The key insight: DeepMind’s agents wrote about 6.5 million lines of machine-checked proof for libsignal’s compiled code, and the paper says the proof work took a team of five four weeks: 20 person-weeks against the twenty person-years it cites for seL4, 52 times fewer, by our arithmetic, for systems that differ in size and scope.
The target was libsignal, the reference implementation of the Signal protocol, in version 0.105.4 with the spqr 1.6.0 crate. The team compiled the unmodified Rust code with the stock rustc compiler to AArch64 machine code and proved properties of that executable. The paper puts the verified code at about 560 KB, about 140,000 AArch64 instructions.
The proofs cover Signal’s post-quantum PQXDH handshake and the Triple Ratchet. The paper says they establish “forward secrecy, authentication, post-compromise security, and handshake secrecy”. It calls them “the first proofs connecting protocol-level security guarantees to the unmodified code of a widely used cryptographic protocol implementation”.
The paper notes that variants of the protocol are deployed in WhatsApp, Meta’s Messenger app and Google Messages, as well as the Signal app. DeepMind says it is open-sourcing its formal verification artifacts.
Human effort on three formal verification projects in person-weeks. The seL4 and CompCert figures are as the AlphaProof Code paper cites them; the Signal figure is the paper’s team of five over four weeks. Conversions to person-weeks are our arithmetic; the systems differ in size and scope.
A Chinese-language post by @VoxVex000 on 10 November, quoting Pushmeet Kohliโs announcement. The poster claims DeepMind produced 6 million lines of Lean proofs in one month using Gemini 4 Carbon and AlphaProof Code, that Google held Gemini 4 Carbon back for two weeks, and that the work adds value to Alphabetโs $567.7 billion of cloud orders. DeepMindโs blog says over six million lines of Lean and four weeks, with agents powered by Gemini Carbon (the paper says about 6.5 million lines); Alphabetโs 10-Q gives $567.7 billion of remaining performance obligations, $560.8 billion of it Google Cloud. The two-week claim and the link to the backlog are the posterโs own; we have not verified the first.
How the Agents Split the Work
AlphaProof Code splits the job in three. A specification-writer agent works with a human to turn models from the research literature into Lean statements, reviewer agents critique each draft, and the human approves it. A team of prover agents then proves the approved statements about the executable.
A gatekeeper agent accepts a proof only if Lean’s kernel checks it. The paper says a proof team “typically had 40 to 80 agents and ran on a single machine for 24 to 72 hours”. It says the proofs “were produced entirely by agent teams powered by Gemini 4 Carbon” and that “none of the authors wrote any proof code”.
The specifications were written through an exchange between the researchers and agents powered by Gemini 4 Carbon and Claude Opus 5.5, the paper says. DeepMind’s blog says the agents audited the proof and flagged “just one percent of the codebase” for a second look by a human.
What the Proofs Caught
To test whether the specifications catch real mistakes, DeepMind says it “reintroduced two real, previously fixed bugs into an isolated version of Signalโs code”. The paper says it proved in Lean that the resulting binaries violate the symbolic security theorems. In one, the paper says, a prekey message that decrypted only under an archived session made libsignal store the new session’s state anyway, including a copy of the signed prekey’s secret.
The paper also reports errors in published analyses of the protocol that had to be repaired before the proofs went through. It says “Fifteen are gaps or bugs in the computational analysis of the Double Ratchet”, and it lists problems in printed ProVerif queries from another analysis.
What Signal Says
Ehren Kret, Signal’s chief technology officer, is quoted in DeepMind’s post: “While prior formal verification work has necessarily focused on smaller scopes, this new approach enables a significant expansion”. He adds: “As AI-driven security threats grow, having the core Signal protocol rigorously verified with automated theorem proving gives us even greater confidence that our code delivers on its privacy guarantees”.
Formal Verification as a Cost Line
The paper frames the result as a cost question. It says seL4 took twenty person-years and the CompCert verified C compiler took six, and that “Such labor costs are untenable for all but the most mission-critical domains”. It says “recent AI advances can drastically reduce these costs”.
By our arithmetic, the agents produced about 250 lines of proof for every line of trusted specification a reader must check, and about 325,000 lines of proof per person-week. The paper’s own lesson is that “once proofs become cheap, the hard question is” what to prove: it says the majority of the team’s effort went into specifications, not proofs.
As we read the paper, the scarce input moves from proof writing to specification: it says the majority of the team’s effort went into specifications, and that no author wrote proof code.
As we read the blog and paper together, the security pitch is a guarantee rather than a bug count: the proofs establish stated properties of one compiled build, under stated assumptions, and the test was whether two known bugs would break them.
Google DeepMind, AlphaProof Code technical paper
“In this paper, we show that recent AI advances can drastically reduce these costs”
Three Implications
FOR SECURITY BUYERS The paper sets formal verification against reactive bug finding, which it says has inherent limitations even with advanced AI.
FOR SOFTWARE TEAMS The trusted part a reader must check is about 26,000 lines, against about 6.5 million lines of agent-written proof, the paper says; that is about 250 to 1, by our arithmetic.
FOR THE MODEL MAKERS The specifications were written with agents powered by Gemini 4 Carbon and Claude Opus 5.5; the proofs came from Gemini 4 Carbon agents alone, the paper says.
The framework argues: “Universal harnesses are likely to dominate highly digital, modular, and verifiable work.”
As we read it, formal proof is verifiable work in its purest form: Lean’s kernel accepts or rejects every step, which is why DeepMind could let agent teams run for days with no author writing proof code. The framework predates this paper and does not discuss it.
What Is Not Established
The guarantees have stated limits. The paper says its symbolic guarantees “assume perfect cryptography”, that its computational bounds “give no guarantee against a quantum adversary”, and that more than two parties and multi-device state are not covered. The results concern the team’s own build of libsignal-cli.
The blog post and the paper sections we read give team time, agent counts and run lengths, but we did not find a dollar figure for the compute. DeepMind’s blog says scaling these techniques “requires continued research”. The paper reports reintroduced, already-fixed bugs; it does not describe a previously unknown bug in the shipped libsignal code in the sections we read.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
DeepMind’s agents wrote about 6.5 million lines of machine-checked proof for libsignal’s compiled code in four weeks of work by five people, the paper says. That is 20 person-weeks against the twenty person-years it cites for seL4, 52 times fewer, by our arithmetic, for systems that differ in size and scope.
A note on sourcing. This piece rests on Google DeepMind’s blog post and technical paper of 10 November 2026, both read in full, and on Pushmeet Kohli’s post on X. Ehren Kret’s statement is as published in DeepMind’s post. Person-week conversions and ratios are our arithmetic. Signal and DeepMind were not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
