04 · Technology · United States

AI Checks Signal's Math

Google DeepMind says AI agents mathematically proved key security guarantees in the code behind the encryption used by Signal, Google Messages and WhatsApp.

A private chat on a phone.

A private chat on a phone.Illustration by The Good Times

In one minute

Google DeepMind said its researchers used a new AI system, AlphaProof Code, to formally verify key security properties of the code that runs the Signal protocol, the encryption behind private chats in the Signal app, Google Messages and WhatsApp. Formal verification means proving with mathematics that software behaves as intended, rather than testing it and hoping. AI agents wrote more than six million lines of proof in Lean, a language that checks proofs automatically, and flagged about 1% of the work for human experts to review. DeepMind says it is the first time the implementation of a protocol this widely deployed has been formally verified by any method. The project took about four weeks. It found no new flaws, and when run on old versions of Signal with known, long-fixed bugs, it caught them. The proof covered core parts of the protocol, including that messages stay protected before and after a device is temporarily compromised, not the whole Signal app. Signal's chief technology officer, Ehren Kret, said it gives the company "even greater confidence" in its privacy guarantees. Google would not say what the work cost.

Google DeepMind researchers used AI agents to mathematically prove key security guarantees in the actual code that runs the Signal protocol, the encryption system that protects billions of conversations in the Signal app, Google Messages and WhatsApp.

The tool, called AlphaProof Code, is an agent-based framework for formal verification, the practice of proving with mathematics that software does exactly what it is supposed to do. Security researchers have long treated full verification of widely used code as a grand challenge, because writing the proofs by hand takes enormous time. DeepMind says this is the first time AI has been used to verify a widely deployed cryptographic protocol at this depth, and the first time the implementation of such a protocol has been formally verified by any method.

How it was done

Researchers first worked with several AI models to turn Signal's security goals into formal mathematical statements, and human experts reviewed every one. Agents running on Google's Gemini models then wrote the proofs, more than six million lines of Lean, a programming language whose checker confirms each step automatically. The agents also audited their own proofs and marked about 1% of the work for a human second look. The project ran as a four-week proof of concept, helped by earlier verification work the Signal Foundation had done.

"We don't want this sort of vibe world where AI is generating the goals and AI is generating the proofs," Swarat Chaudhuri, a DeepMind research scientist, told WIRED. "There's a middle ground."

What it proved and what it did not

The analysis found no new vulnerabilities in the parts of the protocol it covered. DeepMind said the proofs reach the compiled code running on devices, including the guarantee that messages stay protected before and after a device is temporarily compromised. Run against older versions of Signal that contain known, long-fixed flaws, the system identified them and showed those versions could not be verified. It did not verify the entire Signal mobile app, or other parts of the service beyond the core protocol code under review. DeepMind is releasing the specifications and proofs so others can check and build on them.

"Google DeepMind's approach enables significant expansion," said Ehren Kret, Signal's chief technology officer, adding that it gives the company "even greater confidence in our code's privacy guarantees." Pushmeet Kohli, DeepMind's vice president of research, said banks and other companies that worry about bugs could one day prove such properties for their own systems. Google did the work free as research, declined to give its cost, and would not say whether it plans a commercial product. Whether the method can scale to large amounts of software at a workable price is still an open question.

Sources

More from today

02 · Politics · Washington

House Shelves Hegseth Impeachment

The House voted 211-204 on a motion to refer Rep. Thomas Massie's impeachment resolution against Defense Secretary Pete Hegseth to the Judiciary Committee; Massie was the only Republican to vote no.

03 · Housing · New York

East Harlem Buildings Go to Auction

Lenders auctioned the first 15 of 38 rundown rent-stabilized buildings in East Harlem after rejecting a city-backed buyer; only one drew a bid.

05 · Trade · Beijing

China's Surplus Passes $1 Trillion

China sold $1.04 trillion more goods than it bought in the first 10 months of the year, driven by global demand for AI hardware.

08 · Opinion

Waiting on the Free Bus

The war over free buses was lost in Albany months ago. City Hall is still fighting it in press statements, and the mayor owes New Yorkers a plain answer.