The facts

  1. DeepMind says its agents generated formal proofs for selected security properties in Signal’s libsignal code, with Lean checking the proofs against stated specifications. [1][2]

  2. The paper describes properties of PQXDH and Double Ratchet implemented in the command-line/reference version of libsignal; it does not claim to verify the full mobile applications. [2][4]

  3. DeepMind reports that the work took four weeks and that the system detected two deliberately seeded bugs in a test version. [1][2]

  4. The proof framework depends on human-written specifications and modeled platform semantics; the results establish only the properties within that defined scope. [2][3][4]

  5. WIRED reported the exercise did not discover new vulnerabilities in the scoped production code and characterized it as a proof-of-concept rather than a complete Signal-app audit. [4]

Still unconfirmed

  • Whether the approach can scale to the full Signal client and other production systems remains unknown.
  • The guarantees cover only the properties and implementation assumptions specified in the paper; other properties are outside its scope.

Sources

References support the numbered facts above. Source checks: 10 Nov 2026, 23:30 UTC.

The DeepMind announcement and research paper are primary sources for the method and claimed properties. The Register and WIRED independently explain limitations. The story does not imply a complete audit of Signal’s apps or a guarantee that the entire protocol is bug-free.

  1. Introducing AlphaProof Code

    Google DeepMind · Primary source · 10 Nov 2026

    Opened the official announcement, which describes the result, testing setup and Google’s claims about agent performance.

  2. Formal Verification of Signal 4

    Google DeepMind · Primary source · 10 Nov 2026

    Opened the technical paper describing the proof scope, specifications, Lean verification and implementation boundary.

  3. Google DeepMind melds humans and AI agents to vet Signal’s security techniques

    The Register · Reporting · 10 Nov 2026

    Opened independent coverage discussing the proof’s scope, human input and assumptions.

  4. A new Google DeepMind tool formally verified the Signal protocol

    WIRED · Reporting · 10 Nov 2026

    Opened independent reporting that calls the work a proof-of-concept, explains its limits to selected code and says no new vulnerabilities were discovered.

Report a factual error