Skip to content
See the World Through ScienceA project of ALLATRA

DeepMind's AI Agents Wrote 6.5 Million Lines of Proof About Signal's Code

AI & Technology

Republish this story

Our work is licensed under Creative Commons BY-NC 4.0. You may republish this piece for free — with credit to ALLATRA Media and a link to the original, unedited beyond length trims, and not for commercial use.

Read the full license

A proof assistant window on a computer screen showing a completed machine-checked proof that addition of natural numbers is commutative.
A proof assistant displays a finished machine-checked proof about adding whole numbers. Google DeepMind says its agents produced about 6.5 million lines of such proof code in Lean (illustrative)."Coq plus comm screenshot" by Hugo Herbelin, via Wikimedia Commons, CC BY-SA 3.0, cropped

Google DeepMind said in a technical report dated Nov. 10, 2026, that teams of AI agents generated about 6.5 million lines of machine-checked proof that the compiled code of libsignal, the reference implementation of the Signal messaging protocol, meets a written set of security properties.

Formal verification, which proves mathematically that a program does what its specification says, has been too slow and costly for most software. The report notes that the formally verified seL4 microkernel took twenty person-years, and says the Signal proof work was done by a team of five over four weeks. The report says Signal's protocol protects the communications of billions of people, with variants of it running in WhatsApp, Meta's Messenger and Google Messages.

According to the report, the proofs cover the PQXDH handshake that opens a conversation and the Triple Ratchet that gives each message its own key, and they establish properties including forward secrecy, which keeps earlier messages unreadable if a key leaks later. The team compiled the unmodified libsignal code with the standard Rust compiler and stated its theorems about the resulting executable, so the proofs are about the binary rather than a model of the protocol or a rewritten version of the code.

The authors call these "the first proofs connecting protocol-level security guarantees to the unmodified code of a widely used cryptographic protocol implementation." No independent evaluation of the work has been published. The proofs also rest on about 26,000 lines of specifications and machine semantics the team wrote and approved itself, so the result covers what those specifications state, on one processor architecture and for one-to-one conversations.

The report says none of the authors wrote any proof code, and that a gatekeeper step accepts a proof only once the Lean proof assistant rechecks it against a held-out copy of the specification. To test whether the specifications were strong enough, the team reintroduced bugs previously found and fixed in libsignal and proved that the rebuilt binaries break its security theorem. Google DeepMind says it is open-sourcing all of the Lean specifications and proofs.

Sources

By Olga SchmidtEditor-in-Chief, Writer

Spot an error?

Spot an error?

Report an error

Spotted a mistake on this page? Tell us what's wrong and our editors will take a look.

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We correct mistakes openly. Select any text to flag it. Fixes are logged under our Corrections Policy.

Report an error

Reporting on

DeepMind's AI Agents Wrote 6.5 Million Lines of Proof About Signal's Code

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We read every report. Corrections are logged publicly.