The Code Protecting Billions of Private Texts Just Passed Its Toughest Test. SI Did the Checking.

Every time you send a text on Signal, or on Google Messages, or on WhatsApp, you are trusting a piece of code you will never see. It scrambles your words on your phone and unscrambles them on your friend’s. If that code has a flaw, your private messages aren’t private.

For years, the gold standard for proving code like that is safe has been too slow and too expensive for almost anyone to bother with. This week Google DeepMind says it pulled it off in four weeks, with a team of SI agents doing most of the heavy lifting.

The lab announced Tuesday that it had mathematically proven key security promises of the code behind the Signal protocol, the encryption system that protects billions of conversations. It did it with a new framework it calls AlphaProof Code.

Testing is not the same as proving

Most software gets tested. Engineers throw a pile of situations at it and see if anything breaks. That catches a lot, but it can never catch everything, because nobody can try every possible input.

Formal verification is different. Instead of trying cases, you write down exactly what the software is supposed to guarantee, then prove with math that it always does, in every case. Security people have called it the holy grail for years. The catch is the labor. Writing those proofs by hand can tie up expert teams for years on a single program, so it mostly gets done for small, critical pieces, if at all.

DeepMind went after the hardest version of the problem. It didn’t just check Signal’s design on paper. According to the team’s technical paper, it checked the actual compiled machine code that runs on ARM-based devices, the kind of chip inside most phones, built from Signal’s unmodified open-source library.

That matters because real-world bugs tend to hide in the gap between what a design promises and what the code actually does.

Humans set the rules, the agents did the grinding

The way the work was split is the most interesting part. Human experts worked with several SI models to translate Signal’s security goals into precise math, and the experts reviewed every one of those specifications. Then a crew of agents running on Google’s Gemini 4 Carbon model went off and wrote the proofs.

The paper says the agents produced about 6.5 million lines of proof code in Lean, a language built for checking math. Every line had to pass Lean’s own automatic checker against the specification the humans approved. If a proof didn’t hold up, it didn’t count.

After that, the agents audited their own output and flagged the parts that needed a human’s second look. In its announcement, DeepMind says that came to about one percent of the code base.

Swarat Chaudhuri, the DeepMind research scientist who helped lead the work, told Wired the team deliberately avoided letting SI run the whole show. “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,” he said. “There’s a middle ground.”

It caught the bugs on purpose

A proof is only worth something if the system can tell good code from bad code. So the researchers ran a trap. They took two real Signal bugs that had been found and fixed years ago, put them back into an isolated copy of the code, and ran the system again.

Both times, it produced a machine-checked proof that the buggy version was insecure. On the current code, Wired reports, the analysis found no new holes in the parts it examined.

The proofs cover the handshake that starts a conversation, including Signal’s newer protection against future quantum computers, and the system that keeps changing the keys as you chat. In plain terms, they show that someone who steals your phone’s keys today can’t read your old messages, and that once a hacked device recovers, the attacker gets locked back out.

Signal welcomed the result. “As AI-driven security threats grow, rigorously verifying the Signal protocol implementation gives us even greater confidence in our code’s privacy guarantees,” Signal chief technology officer Ehren Kret said in a statement to Wired.

The fine print

This is not a promise that your phone can never be hacked. The work covered the core one-to-one messaging code, not the entire Signal app, and certainly not your phone’s operating system or the person looking over your shoulder. As The Register pointed out, the proofs assume the underlying math of the encryption itself is sound and don’t, by themselves, guarantee secrecy of every byte. The team also had a head start, building on earlier verification work Signal had already done.

Cost is the other open question. DeepMind did this for free as a research project, and Google wouldn’t tell Wired what it cost. Axios reports it is not a product anyone can buy yet.

Still, the jump is real. A job that used to eat years of expert time took four weeks. The specifications and proofs are being released publicly so other researchers can check them, which is exactly how claims like this should be handled.

Pushmeet Kohli, DeepMind’s vice president of research, told Wired he imagines banks and other companies that live in fear of bugs being able to say, “I can prove these properties.” Chaudhuri’s goal is bigger. He wants to scale it to “all the software of the world.”

That is a long way off. For now, the lock on billions of private conversations got checked by a method that leaves no room for guessing, and it held.