Google DeepMind Says Its AI Checked Signal's Encryption Code in Four Weeks: What That Does and Does Not Prove

What DeepMind announced

On 10 November 2026, Google DeepMind introduced AlphaProof Code, which it calls an agentic framework for formal verification. Formal verification means writing a mathematical proof that software meets a precise rule, then having a program check that proof. DeepMind says it used the approach on the code behind the Signal messaging protocol, the code that runs on people's devices.

DeepMind says the work took about four weeks and notes that expert teams have historically needed years for comparable projects. Those timings are the company's own.

What was checked

The properties checked are about keeping messages protected. One example DeepMind gives is that messages stay protected before and after a device is briefly compromised. The proofs reach down to the compiled code, not just the written design, which is where earlier efforts mostly stopped, according to press coverage of the announcement.

DeepMind says the AI agents were built on its Gemini Carbon model and wrote more than six million lines of proof in a language called Lean. Its own blog page also shows a meta description that says three million, and the page does not explain the gap, so treat the exact figure with care.

Who did what

• Human experts chose what to prove and, DeepMind says, reviewed every specification.
• The agents wrote the proofs and audited them, flagging about 1 percent of the codebase for a human second look.
• Lean, the proof checker, validates the rest automatically, so the result does not rest on trusting the AI's word.

The test DeepMind describes

To show the system can catch real mistakes, DeepMind says it put two previously fixed bugs back into an isolated copy of Signal's code. In both cases the system produced a machine-checked proof that the buggy version was insecure. The sources we read describe no other validation of the system.

What this does and does not prove

A proof is only as good as the rules it checks. It shows the code meets the properties experts wrote down; it does not show the code is safe against every threat, such as a stolen phone or a flaw in a property nobody thought to specify. DeepMind itself says scaling these methods needs more research, and the press report we read does not address whether the specifications were independently checked.

Signal's chief technology officer, Ehren Kret, praised the approach as a major breakthrough, according to DeepMind's page. DeepMind also describes this as the first AI-driven verification of a widely used cryptographic protocol at this depth. That is a claim from the company, and we could not confirm it independently.

What it means for ordinary users

Nothing changes in the Signal app today, and you do not need to do anything. The wider point is that AI is starting to be used to check security software, not only to write it. DeepMind says it is releasing its verification files as open source and plans to extend the method to critical systems such as power grids and medical devices. We have not tested any of this ourselves.