DeepMind Formally Verifies Key Guarantees of the Signal Protocol
Google DeepMind used a new AI system, AlphaProof Code, to demonstrate several security guarantees of the Signal protocol. The team found no new flaws in the studied parts, but the method remains limited to a specific scope.
Google DeepMind used AlphaProof Code, an AI system dedicated to mathematical proofs, to verify several security guarantees of the Signal protocol, the cryptographic layer that also protects WhatsApp and Google Messages. According to Wired, the study lasted four weeks and revealed no new vulnerabilities in the examined parts. The stakes go beyond Signal: formal verification promises to prove that a specific bug does not exist, rather than just searching for it. Not exactly the kind of conclusion that makes you throw your hands up in the air, but it is a serious advance.
A mathematical proof, not just another test
Formal verification consists of demonstrating through mathematics that software indeed respects its initial properties. In theory, this avoids relying solely on tests which can miss a defect. In practice, the exercise remains slow, as properties must be translated into formulas, then the proof constructed. DeepMind explains that its agents automate a large part of this last step, while leaving humans to control the key points.
The scope studied remains narrow: AlphaProof Code did not prove all of Signal, but only certain cryptographic guarantees of the protocol. DeepMind also tested the tool on older versions known to contain vulnerabilities that have already been fixed. Result: the system correctly flagged them and concluded that these versions could not be verified, which corresponds to the expected behavior of a reliable verifier.
Signal serves as a testbed, but the method aims wider
Signal did not fall into the blind spot of formal verification by chance. The app has already integrated this type of work into the design of its post-quantum ratchet, with the help of Cryspen, to mechanically verify its protocols. Ehren Kret, chief technology officer of Signal, stated: « While prior formal verification focused on smaller portions of the code, Google DeepMind’s approach enables significant expansion. As AI-driven security threats grow, rigorously verifying the Signal protocol implementation gives us even greater confidence in our code’s privacy guarantees. »
On DeepMind’s side, Swarat Chaudhuri, research scientist, summarized the project’s philosophy as follows: « 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. There’s a middle ground. » This nuance matters, as full automation remains far from achieved. The researcher also acknowledged there is « a long distance to go » before aiming at « all the software of the world ».
What Google does not yet say about the cost
The work was carried out free of charge, as part of a research project, and Google did not indicate what it would cost if it had to be industrialized. The company also did not say whether it intended to sell formal verification as a product or service. This is often where the fine technical promises meet the price tag, and the roadmap.
For DeepMind, the strategic interest is clear: if AI can help certify critical cryptographic building blocks, the approach could interest banks, software publishers, or services handling sensitive data. Pushmeet Kohli, vice president of research, put it this way: « If you can now verify the correctness of these implementations, then you can rest easy. You can imagine banks, imagine all those other companies who rely on assurances and who worry about bugs who could now say, ‘I can prove these properties.’ »
A cybersecurity race broader than Signal
This announcement comes as several AI labs try to apply their models to security. OpenAI says GPT-6 Cyber found 81 hidden flaws in Firefox; Google gave cybersecurity defenders early access to Gemini 4 Carbon; Anthropic claims Claude solved a 25-year-old NSA cryptographic puzzle. In this context, DeepMind bets on a less spectacular but more robust approach: not just hunting bugs, but proving that part of the code cannot contain certain ones.
The real question remains, the only one that ultimately matters for the market: can this method be extended to more complex software, at large scale, and at a reasonable cost? For now, DeepMind provides a credible demonstration on a protocol already well studied. The leap to the rest of the world’s software is not yet written.
By the numbers
- 4 weeks — duration of the research project, according to Wired.
- 3 applications — Signal, WhatsApp, and Google Messages use the same cryptography.
- 81 flaws — hidden bugs found by GPT-6 Cyber in Firefox, according to OpenAI.
- 25 years — age of the NSA cryptographic puzzle solved by Claude, according to Anthropic.
- 9 problems out of 353 — open Erdős problems tackled by AlphaProof Nexus, according to arXiv.