DeepMind’s AlphaProof Code writes 3 million Lean proof lines to verify Signal encryption

DeepMind’s AlphaProof Code writes 3 million Lean proof lines to verify Signal encryption

On November 10, 2026, DeepMind introduced AlphaProof Code, an agentic system that generated three million lines of Lean proofs to verify the core encryption of Signal. The result moves AI formal reasoning from competition mathematics into production security software, where machine-checkable proofs can supplement or replace parts of manual cryptography audits. Liu Gong's take is that the bottleneck in formal verification is shifting from human proof labor to reproducible engineering pipelines,

DeepMind published the blog post at 12:02 on November 10, 2026 under the title Introducing AlphaProof Code: a breakthrough in verifying complex software at depth. The post's central claim is that multiple AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal's core encryption. Signal is known for end-to-end encrypted messaging, so a logical gap in its core encryption would affect message confidentiality and identity authentication. The system turns code behavior, protocol assumptions and security goals into theorems that Lean can check, then searches for proof scripts; the three million proof lines count only after they pass the Lean kernel. DeepMind calls the work a breakthrough in verifying complex software at depth. The public post does not disclose model size, training cost, a complete proof repository address or an independent audit result. For

Formal verification is not new. Lean was initiated by Leonardo de Moura at Microsoft Research and has become an important infrastructure for mathematics and program verification over the past decade; Signal has long been treated as a high-security communications benchmark. In the past, verifying a core protocol required experts to translate protocol rules, cryptographic assumptions and implementation details into theorems, which made cost and cycle time high. Most products therefore rely on code audits, fuzzing and penetration tests, which cannot cover every state. AlphaProof previously showed that AI could search formal proofs in competition mathematics; its 28/42 score at the 2024 IMO

Liu Gong's take: AlphaProof Code's three million Lean proof lines are not a one-off demo but a turning point from handcrafted formal verification to agentic proof pipelines; within 12 months, the first changes will appear not in consumer apps but in security acceptance for high-assurance communications, automotive chips and industrial control.
From a technical standpoint, the value lies in reproducibility. If one expert can review 1,000 lines per day, three million lines equal 3,000 person-days, or about 8.2 person-years, before theorem modeling and defect localization. Once proofs are written as Lean scripts, reviewers do not have to read a natural-language report line by line; the Lean kernel can replay the checks. This differs from penetration testing: sampling raises the probability of finding bugs, while formal proof attempts to cover all states described by the theorem. The boundary is equally clear. Proofs are valid only for the modeled scope; implementation bugs, compiler behavior and hardware side channels remain outside unless explicitly modeled. That is why the choice of Signal's core encryption, rather than sample code, matters: the closer the target is to a real attack surface, the harder it becomes to avoid theorem modeling and scope boundaries.
On the business side, DeepMind is moving AlphaProof from math contests to software verification, so the selling point shifts from model capability to reproducible proofs and audit integration. Certification bodies, telecom vendors and chip companies need versioned proof repositories, failed-lemma lists and regression toolchains, not a blog post. Within the next six months, Liu Gong predicts two kinds of procurement pilots: one using formal proofs as a release gate for encryption protocol upgrades, another using them for regression verification of secure boot modules. Whoever packages proof generation, checking and defect localization into a service will capture the first budget. Conversely, if the proof repository stays closed and the theorem list opaque, customers can only treat the work as publicity and cannot put it into procurement comparisons.
For the industry, the division of security audit work will be redrawn. Human experts still define threat models and acceptance theorems; agents handle large-scale lemma search and proof drafts; audit bodies review proof repositories and toolchains. In AI visual inspection, Liu Gong has seen that deployment depends on whether acceptance criteria can be reproduced, not on how impressive a single demo looks. Industrial quality inspection relies on traceable samples, thresholds and regression sets; formal verification relies on traceable theorems, proofs and versions. If three million proof lines are treated as a marketing number, the industry will repeat the old pattern of bright demos and hard maintenance. If they become part of continuous integration, the release cadence of security software will be rewritten.

  • Three million Lean proof lines equal 3,000 person-days, or about 8.2 person-years, if one expert reviews 1,000 lines per day; theorem modeling and defect localization are extra.
  • AlphaProof scored 28/42 at the 2024 IMO and moved to Signal core encryption verification in 2026, a gap of about 28 months from competition math to industrial software.
  • At an estimated 60 characters per line, three million lines amount to roughly 180 million characters; at one proof step per line, they represent three million machine-checkable checkpoints.
  • The target is Signal's core encryption rather than sample code; if the scope covers protocol and implementation, it exceeds routine code audit, but the

For the industry, within three months the DeepMind post will push formal-verification teams to reproduce the experiment and decompose the theorem structure, while security vendors add machine-checkable proofs to proposals; independent third-party audit will not be completed immediately, and the proof scope for Signal's core encryption still needs line-by-line review. Within one year, if proof repositories and regression tooling mature, encryption protocols, automotive chips and industrial control may put formal proofs into release gates, shifting security audit from reports to proofs. For ordinary users, Signal will not feel faster or slower in the short term; the change is that encryption

After the DeepMind post, no citable formal response from competitors was available in public information; discussion centered on three points. Competitors will focus on reproducibility: whether the three million proof lines come with a public repository and can be replayed on comparable hardware will determine whether this is a purchasable audit product. Formal-verification experts will ask about scope: which protocol assumptions, implementation details and cryptographic primitives of Signal's core encryption were written into theorems, and which were left outside the model. The developer community is more focused on tooling: whether Lean checks can be embedded in CI and whether failed

Questions worth exploring:

  • How formal proofs can be integrated into CI/CD release gates
  • Reproducibility of three million proof lines and third-party audit standards
  • How AI proof agents reshape security engineering roles and the audit industry

Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.

This item comes from DeepMind Blog (published 2026-11-10 12:02:28). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.


Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private

Leave a Reply

Your email address will not be published. Required fields are marked *

中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn