Signal’s logo and wordmark, from Signal’s brand assets. Image: Signal
Google DeepMind has used a new AI system it calls AlphaProof Code to mathematically prove several of the Signal protocol’s key security guarantees, the cryptography that also protects chats on WhatsApp and Google Messages, Wired reported. The four-week research project found no new vulnerabilities in the parts it examined, and DeepMind is releasing its specification and proofs.
What DeepMind proved, and what it didn’t
Formal verification uses maths to show that software does what it is meant to do, rather than testing it and hoping nothing slips through. Security researchers have long called it a holy grail, but it has been painfully slow: people have to describe the properties they want, turn them into formulas, then build the proof by hand. DeepMind’s agents take on much of that last step.
The work covered selected cryptographic guarantees of the Signal protocol, not the whole Signal app. As a check, the researchers ran AlphaProof Code on older versions of the protocol that contained known, long-fixed vulnerabilities. It correctly flagged them and showed those versions could not be verified, which is what a working checker should do.
DeepMind did the analysis for free as a research project. Google declined to say what it cost, or whether it plans to sell formal verification as a product or service.
“There’s a middle ground”
The researchers say their biggest advance is a process in which AI agents contribute and correct themselves at almost every stage while people review the work at key points. Swarat Chaudhuri, a DeepMind research scientist, told Wired the goals should not come from the AI alone:
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.
Swarat Chaudhuri, research scientist, Google DeepMind
He also played down how far it goes for now: there is “a long distance to go” before the method scales to “all the software of the world”, he said, though that is the aim.
Signal gave DeepMind a head start
The team says Signal’s own earlier verification work helped it move quickly. Signal has made formal verification part of designing its post-quantum ratchet, working with the firm Cryspen to machine-check its protocols. Ehren Kret, Signal’s chief technology officer, said in a statement:
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.
Ehren Kret, chief technology officer, Signal
From maths olympiads to real code
The name ties the tool to AlphaProof, the DeepMind system that writes proofs in the Lean language and, with AlphaGeometry 2, reached silver-medal standard at the 2024 International Mathematical Olympiad, described in Nature last year. In May, DeepMind researchers reported that an agent called AlphaProof Nexus had settled 9 of 353 open Erdős problems for a few hundred dollars each. DeepMind hasn’t said how AlphaProof Code relates to either system.
Pointing that kind of rigour at code that billions of messages depend on is a different test. Pushmeet Kohli, DeepMind’s vice president of research, pitched it at firms that worry about bugs:
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.’
Pushmeet Kohli, vice president of research, Google DeepMind
It arrives as AI labs race to aim their models at security. OpenAI says GPT-6 Cyber found 81 hidden Firefox flaws, Google gave cyber defenders first access to Gemini 4 Carbon, and Anthropic says Claude cracked a 25-year NSA cryptography mystery.
Why it matters
Most AI security work so far finds bugs; this proves that, for the parts checked, certain bugs cannot exist. Whether that scales beyond a single, already well-studied protocol, and at what price, is the question DeepMind hasn’t answered yet.
Sources: Wired, Signal, Nature, arXiv (AlphaProof Nexus)
Latest Google news
