DeepMind verifies the Signal protocol with AI agents and Lean
AlphaProof Code proved forward secrecy, authentication and post-compromise security of libsignal in Lean, at machine-code level: limits and impact.

On 10 November 2026 Google DeepMind announced that it had formally verified some of the main cryptographic guarantees of libsignal, the reference implementation of the Signal protocol, using a new system called AlphaProof Code. The proven properties concern the post-quantum PQXDH handshake and the Triple Ratchet: forward secrecy, authentication, post-compromise security and handshake secrecy. Versions of the Signal protocol are also used by WhatsApp and Google Messages.
The analysis found no new vulnerabilities in the parts it examined. There is therefore no patch to install: the news matters for what it says about the robustness of the protocol and about how AI can make formal methods practical.
What happened
According to The Register and SecurityWeek, a team of five researchers completed the work in four weeks. The details are in the paper “AlphaProof Code: Verifying Signal’s Security at the Machine-Code Level”. Google is publishing the specifications and proofs so that others can reuse them.
Wired, whose article was republished by Ars Technica, reports that DeepMind carried out the analysis free of charge as a research project, without disclosing its cost and without commenting on a possible commercial product. The researchers acknowledge a head start thanks to formal verification work previously done by Signal. Signal’s CTO, Ehren Kret, commented that earlier work covered smaller portions of the code and that this approach allows a significant expansion.
The Register links the result to the challenge issued in April by the Beneficial AI Foundation to verify the Signal protocol and app with Lean: it is a step in that direction, not the finish line.
How AlphaProof Code works
AlphaProof Code is a multi-agent framework for writing formal specifications and proving them, with people involved at the key steps. The phases, reconstructed by cross-checking the sources:
- Specification: Signal’s security goals are translated into a mathematical specification in Lean, reviewed by people. According to SecurityWeek, the researchers worked with agents based on Gemini 4 Carbon and on Claude Opus 5.5.
- Machine-code semantics: the system models the behaviour of the compiled code, not just the protocol design.
- Proof: agents based on Gemini Carbon generated around 6.5 million lines of Lean proofs; the Lean kernel checks every step against the specification and rejects invalid proofs.
- Review: an audit pass flags the parts that need human examination (The Register).
The most relevant technical point, stressed by SecurityWeek, is that the guarantees hold for the compiled machine code, while most protocol verification stops at the design. This narrows the gap between the proven model and the program that actually runs.
To show that the method catches real flaws, the team applied AlphaProof Code to builds containing two old libsignal vulnerabilities, fixed long ago: in those cases the proof does not go through and the security properties are shown to be broken.
For scale, the paper quoted by The Register notes that formally verifying the seL4 microkernel took 20 person-years and developing the verified CompCert compiler took six years.
The stated limits
The sources agree: the result is important but bounded.
- Ideal cryptography: the proof assumes perfect cryptographic primitives and does not guarantee the computational secrecy of bytes (The Register).
- Not the app: according to SecurityWeek the results cover Google’s own build of the code, run through a small command-line wrapper, and “not Signal’s own applications”; Wired also states that the mobile app was not verified in full.
- Side channels: excluded from parts of the analysis (SecurityWeek).
- Cost and scale: still open questions. Researcher Swarat Chaudhuri notes that the next frontier is not the proof but the specification: deciding what to prove remains work that needs people.
The context: Gemini 4 Carbon
On the same day, SecurityWeek reports, Google opened Gemini 4 Carbon to paying developers and businesses through the Gemini API, while the cybersecurity variant remains restricted to members of the Fairwind programme. According to Google, the model has also been used to find and confirm real vulnerabilities. The same kind of tool therefore serves both attack and defence: this is the argument Chaudhuri uses for the urgency of verifying code already in use.
Impact for European and Italian companies
For those using Signal, WhatsApp or Google Messages, the news is reassuring about the core of the protocol, but it does not change day-to-day risk: the concrete threats to business messaging mostly come through compromised devices, outdated apps, devices linked to the account and social engineering, all outside the scope of the proof.
For NIS2 entities, formal verification is useful evidence when assessing supply chain security and secure development (Article 21), provided its assumptions and exclusions are read carefully. The fact that the post-quantum handshake was also verified is worth noting in migration plans towards post-quantum cryptography.
What to do
- No urgent action: there are no new vulnerabilities or patches tied to this news.
- Keep messaging apps updated: the old libsignal flaws used as a counter-test are a reminder that the guarantees only hold for fixed versions.
- Do not stretch the guarantee beyond its scope: a verified protocol does not mean a secure device. Device management (MDM), control of devices linked to accounts and anti-phishing training are still needed.
- If you develop critical software, consider formal methods for cryptographic and authentication components: the specifications and proofs published by Google are a concrete reference.
- In procurement, when a vendor claims formal verification, ask what was proven, on which code and under which assumptions.
- With AI-generated code and proofs, apply the same rule DeepMind followed: a person reviews the specifications, an independent tool checks the results.
How ARXDOME helps
On computers with ARX Agent, the ARX Sentry inventory lists installed software with its version, so you can find desktop messaging clients that need updating. ARX Shield shows the open vulnerabilities on each machine, highlighting those already exploited.
Last updated:


