On November 10, 2026, Google DeepMind published a study reporting formal proofs of selected security properties for compiled libsignal code. The work covers specified Signal operations in one-to-one sessions, not the entire Signal app; its technical paper spells out the code target and the assumptions behind the results.

A formal proof answers a bounded question: whether specified code meets specified properties under a stated model.

What Google DeepMind verified in Signal

The target was libsignal v0.105.4 with SPQR v1.6.0, compiled from unmodified Rust crates into an AArch64 Linux executable using a small command-line harness. The verified code covered the PQXDH handshake and Triple Ratchet messaging for one-to-one sessions. The harness included init, send and recv; keygen and group messaging were outside the stated scope.

That makes this an implementation-level result: the analysis connected security properties to a particular compiled program, rather than considering only an abstract description of the protocol. It does not certify every part of the Signal mobile app or every way the protocol is used.

How AlphaProof Code checks its proofs

A three-stage process moves from a human-approved specification to AI-generated proofs and checks by Lean’s kernel.

Human experts approve formal specifications that state the properties to be proved. AI agents generate proof code, and Lean’s kernel checks whether the proofs follow from those specifications. The agents produce candidate proofs; the kernel, not the agents’ own assertion that a proof is correct, decides whether Lean accepts them.

The paper reports about 6.5 million lines of machine-checked Lean proof code, alongside about 26,000 lines of trusted specifications and AArch64 semantics. Google DeepMind says the project took four weeks and that it open-sourced the Lean development and verification artifacts in its announcement.

Which security properties the study reports

The paper describes two kinds of results. Its symbolic analysis reports ten theorems, including properties such as forward secrecy, post-compromise security, authentication and handshake secrecy. Symbolic analysis represents cryptographic operations through an idealized model, and these results carry the paper’s stated conditions, including fresh randomness and collision qualifications.

The computational analysis gives guarantees for message confidentiality and authenticity, along with forward secrecy and post-compromise security, against bounded classical adversaries under specified cryptographic assumptions. In this setting, the claims concern the adversary’s bounded advantage within the model, not every possible attack.

As a regression check, the researchers reintroduced two previously fixed libsignal bugs. The resulting binaries violated the symbolic security theorems for those cases.

Where the reported guarantees stop

Lean’s checks do not remove every trust assumption. The results rely on the approved specifications, Lean’s kernel, the formal AArch64 semantics, contracts for external routines and the cryptographic assumptions in the paper. That AArch64 model covers 467 instructions and 31 system calls; it excludes threads, signals, the vDSO, dynamic linking and files outside the current directory.

The computational proof covers classical adversaries, not quantum attackers. A symbolic result addresses handshake secrecy after a later Diffie–Hellman break within its specified model, but the paper does not prove SPQR’s post-quantum healing against an attacker that breaks Diffie–Hellman.