deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

DeepMind AI agents reportedly produce machine-checked Lean proof of Signal's core encryption

Google DeepMind says AI agents produced a machine-checked Lean proof of libsignal's post-quantum handshake and Triple Ratchet in four weeks, establishing four security properties for one-to-one sessions.

DeepMind AI agents reportedly produce machine-checked Lean proof of Signal's core encryption

What happened

On November 10, Google DeepMind announced that AI agents had produced a machine-checked formal proof covering the core of the Signal protocol, according to a detailed write-up on dev.to. The work used a framework the post calls AlphaProof Code, powered by an engine named Gemini Carbon, and targeted libsignal, the open-source library implementing Signal's encryption. Specifically, it covered the post-quantum PQXDH handshake and the Triple Ratchet, the two components that protect a one-to-one conversation from the initial key exchange onward.

The proof was written in Lean and runs to roughly 6.5 million lines. Crucially, humans only needed to read about 26,000 of them: around 18,000 lines of specification theorems and 8,000 lines of machine semantics. The rest was machine-generated and machine-checked, which is the entire point of a proof assistant — the checker validates every step, not a person.

The effort reportedly took four weeks. For comparison, the post cites two landmark verification projects: the seL4 microkernel, which consumed roughly twenty person-years, and CompCert, the verified C compiler, which took six. Cutting that timeline to under a month is the headline result.

What the proof establishes

The verification establishes four properties for one-to-one sessions: forward secrecy, meaning leaked present-day keys do not expose past messages; authentication, meaning the handshake binds the session to the intended peer; post-compromise security, meaning a conversation heals after a temporary device compromise; and handshake secrecy, meaning the handshake itself does not leak the session key.

Signal's CTO, Ehren Kret, said the work gives the team "even greater confidence" in the code's privacy guarantees, and WIRED, as cited by the post, reported that the analysis surfaced no new vulnerabilities in the code it examined. A proof that finds nothing is still a result, because it rules out entire classes of bugs in the covered components.

What was actually verified

A detail the post says most coverage gets wrong: the proof concerns compiled AArch64 machine code built from the actual Rust crates — libsignal v0.105.4 and SPQR v1.6.0 — rather than an abstract protocol model. Most cryptographic verification works on a specification, and the gap between that model and the shipped code is where real-world bugs have historically lived. Verifying the binary closes that gap.

The work was also not fully autonomous. Humans wrote the formal specification, decided what "secure" should mean, and kept final approval; the agents did the exhaustive labor of producing millions of lines of proof.

What it does not cover

The limits are stated openly, and the post attributes them to DeepMind's own researchers. Group chats are excluded, as are the Signal application as a whole and the operating system and phone underneath it. Key generation is not verified: the proof assumes the starting keys are sound, so a broken random number generator voids the guarantees.

Most notably for post-quantum observers, the paper reportedly states that no post-quantum guarantee is established for the healing property of SPQR, the Sparse Post Quantum Ratchet that periodically re-injects ML-KEM-768 to keep conversations healing in a post-quantum sense. Only the classical healing of the Triple Ratchet is proved. The symbolic argument also assumes perfect cryptography — that underlying primitives behave exactly as specified — while a separate computational argument bounds what an attacker can do inside the researchers' model.

The accurate summary, the post argues, is not "AI proved Signal secure," but that libsignal's post-quantum handshake and Triple Ratchet satisfy four named properties at the compiled-code level, under stated assumptions. Less exciting, more useful.

Why it matters

libsignal is not only Signal's library: variants of the protocol reportedly run in WhatsApp, Messenger and Google Messages, so a proof of this code reaches well beyond one app. The technique is also public — the specifications and proofs were published, and Gemini Carbon became generally available the same day.

The deeper shift is economic. Formal verification of cryptographic code has historically demanded multi-year, well-funded efforts. A four-week timeline moves it from impossible for a small team to expensive but conceivable, even though the hard human work of writing a correct specification remains. DeepMind's researchers say there is "a long distance to go" before this scales to software generally, and the open questions — group chats, key generation, proving SPQR's post-quantum healing, and bringing the cost down — are the ones to watch. If even a couple fall within a year, the bar for what "we verified our crypto" means will move quickly.

  • #formal-verification
  • #signal-protocol
  • #encryption
  • #lean
  • #google-deepmind
  • #post-quantum

Related posts