DeepMind Mathematically Proves the Signal Protocol Is Secure Using AI

In a four-week project, Google DeepMind researchers applied a new agentic AI framework called AlphaProof Code to formally verify key cryptographic guarantees of the communication protocol used by Signal, Google Messages, and WhatsApp. The work, presented as a research proof-of-concept, found no new vulnerabilities in the portions analyzed — but did something that previously would have required months of manual expert effort: it generated complete mathematical proofs using AI.

What drew attention was not just the result (the protocol is secure), but the method. Traditional formal verification — considered the "holy grail" of software security — relies on engineers writing manual logical specifications and then building formal proofs, piece by piece, to prove that a system behaves exactly as designed. This process historically consumed months or years of specialized expertise. AlphaProof Code automated the translation of high-level descriptions of security properties into low-level mathematical formulas, enabling the AI system to generate proofs that had previously been exclusively human work.

The protocol, the tests, and what was verified

DeepMind researchers did not analyze Signal's entire source code — which would be impractical even for such a system. Instead, they focused on a subset of the mobile app's cryptographic code, precisely the layers handling key establishment and end-to-end encryption. The system was fed a high-level description of the security properties expected of the protocol, and then automatically generated the mathematical formulas required for verification.

When researchers applied AlphaProof Code to older versions of the protocol — those containing vulnerabilities that had already been fixed — the system correctly identified the problems and demonstrated that the protocol could not be formally verified in those versions. For current versions, the system confirmed that the cryptographic guarantees remain intact.

The result is not surprising in the sense that the Signal Protocol (also known as the Axolotl Tree-KEM) has long been widely considered secure. But demonstrating that an AI can prove this mathematically, rather than merely find flaws, is qualitatively different. Formal verification eliminates entire classes of bugs — it does not simply find the ones that exist, but proves that certain categories of error are impossible within the specified domain.

Human-AI collaboration as a model

A frequently overlooked point in the announcement was the role of the collaborative process. DeepMind researchers describe the work as a "human-AI collaboration process," in which agents propose proofs and self-correct, but humans review and correct at critical junctures. Swarat Chaudhuri, a research scientist at DeepMind, was explicit about the limits of this model:

> "We don't want this sort of vibe world where AI is generating the goals and AI is generating the proofs and basically taking tests that are self-written."

This caution is not mere institutional rhetoric. Formal verification depends entirely on the quality of the specification — if you specify poorly, you will correctly prove only what you intended, not what you need. The human role remains crucial for formulating the correct properties to verify, interpreting proof failures, and deciding when the system is reliable enough to proceed.

What Signal and the security community say

Ehren Kret, CTO of the Signal Foundation, noted that prior formal verification efforts covered only small fragments of code, while DeepMind's approach "enables significant expansion" of what can be checked. The Signal Foundation had already conducted some formal verification work on earlier protocol versions, which helped DeepMind researchers get a head start — previous work served as a starting point for the specifications.

The DeepMind team released the formal specification and proof artifacts for public use, allowing other research groups to build on the work. This is particularly relevant given that the Signal Protocol is not used only by Signal: it also protects communications in Google Messages (privacy modes) and WhatsApp, meaning the verification has implications far beyond a single application.

Why this matters for software security

Formal verification has long been considered the highest level of software assurance. Tests can find bugs; formal proofs eliminate entire classes of them. In the context of cryptography, where a single implementation error can compromise years of cryptographic work, the ability to generate proofs automatically is transformative.

The practical question that remains is scalability. DeepMind conducted this project free of charge as research, and Google declined to disclose the cost or announce plans for a commercial formal verification product. Scaling this approach to entire applications, or to software supply chains, remains a challenge — but the publicly released artifacts give the ecosystem a new building block to reach that goal.

Signal's cryptography now has an extra layer of mathematical verification, generated by AI. The protocol has not changed — confidence in it has.

Sources: WIRED, AI Understanding, The Back Room Tech

✓ Independent sources cross-checked and verified before publishing