Google DeepMind researchers have used an agentic AI system to formally verify that substantial parts of the Signal protocol do what they are supposed to do — a machine-checked mathematical proof, not a test suite. WIRED reported the result on 10 November.

Formal verification is the discipline of writing a program's required properties as mathematical statements and then proving the implementation satisfies them, with a proof checker deciding whether the proof is valid. For decades it has been reserved for small, high-value kernels — a crypto primitive, a scheduler — because turning specifications into formulas and formulas into proofs was slow human work.

The Signal analysis took four weeks and used a new framework called AlphaProof Code. Three things stand out.

First, it found no new vulnerabilities in the portions of the protocol it examined. Second, it is not a rubber stamp: pointed at older versions of Signal that contain known, long-fixed vulnerabilities, AlphaProof Code identified them and showed that the protocol could not be formally verified. That negative result is the test that separates a verifier from a demo. Third, DeepMind is releasing the specification and the proof it developed, so others can build on it.

The scope was limited. The researchers did not formally verify the entire Signal mobile app, only parts of the protocol, and they had a head start: the Signal Foundation had already done formal verification of smaller portions of the code.

"While prior formal verification focused on smaller portions of the code, Google DeepMind's approach enables significant expansion," Signal's chief technology officer Ehren Kret said in a statement. "As AI-driven security threats grow, rigorously verifying the Signal protocol implementation gives us even greater confidence in our code's privacy guarantees."

Humans were not removed from the loop. The researchers describe a "human-AI collaboration process" in which agents contribute and self-correct at almost every phase while people review at key junctures — an explicit answer to the obvious objection that a model grading its own proofs proves nothing much.

"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," said Swarat Chaudhuri, a DeepMind research scientist. "There's a middle ground." He added: "We are very excited by this result, but there is a long distance to go before we are able to scale this to all the software of the world, and we want to do that."

The commercial pitch is straightforward. "If you can now verify the correctness of these implementations, then you can rest easy," said Pushmeet Kohli, chief scientist at Google Cloud and DeepMind's vice president of research. "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.'"

The open question is economics. DeepMind ran the Signal analysis free of charge as a research project; Google declined to share the cost and declined to comment on whether it plans a commercial formal-verification product or service. Whether proofs of this kind can be produced at a workable price outside a flagship demonstration decides whether formal verification becomes ordinary engineering practice or remains a spectacle.

One detail matters for anyone reading the headline as an audit: verifying a protocol is not the same as verifying the app that implements it, and a proof is only as good as the specification it starts from. DeepMind's contribution is a credible step on both counts — a specification written down and published, and a machine that can check it — not a certificate that Signal is, in its entirety, secure.