# Google DeepMind’s Lean Proofs Reject Two Bug-Modified libsignal Builds

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/59260
Published: 2026-11-17T10:25:56.000Z
Updated: 2026-11-17T10:25:56.000Z
Section: Technology

> The verification covers a narrow AArch64 Linux command-line target. A Nov. 11 update expanded machine-code checks to 10 routines and reduced trusted hooks from 37 to 27.

Google DeepMind researchers reintroduced two previously fixed bugs in separate libsignal builds, and Lean proofs showed both violated symbolic security theorems, demonstrating that the checks rejected known flaws in the tested program.

AI agents generated about 6.5 million lines of machine-checked proof code for a compiled program using libsignal, the software implementation behind the Signal protocol. Five people guided the four-week project by choosing reference models and reviewing the specifications the agents needed to prove. The specifications express expected security behavior as conditions a proof checker can evaluate.

The verified target was a small command-line program for AArch64 Linux, using libsignal-protocol v0.105.4 and SPQR v1.6.0. The modeled operations were session creation, message sending and message receipt. The theorems linked those operations to security properties for one-to-one messaging through PQXDH and Triple Ratchet.

The verified security goals span authentication, forward secrecy and recovery after a compromise. They address whether older messages remain secret after key compromise, whether a sender is authenticated and whether a session can regain security after a breach. The reported bug checks challenged the symbolic theorems specifically.

The project also uses a computational proof model. Its computational proof covers classical attackers under stated assumptions, while the symbolic model treats cryptography as idealized. The modeled setting excludes multi-device state. Key generation, group messaging and Signal’s phone apps also fall outside the verified scope.

The proof chain still relies on Lean’s checking kernel, reviewed specifications, an AArch64 Linux model and contracts for selected routines. These contracts, called hooks, replace running some routines with assumed descriptions of their behavior. Fewer hooks mean less of the result depends on those modeled descriptions.

A Nov. 11 update added verification for 10 library routines and lowered trusted hooks from 37 to 27. The updated program checks machine code for those routines instead of relying on modeled contracts. The bug checks were part of [Google DeepMind’s Signal verification project](<https://www.tokenpost.com/news/technology/54207>).



## Links in this article

- [Google DeepMind’s Signal verification project](https://www.tokenpost.com/news/technology/54207)
