DeepMind proved Signal's code secure with AI in four weeks: Reveneau

Why it mattersFormal verification on live cryptographic code used to take years. If the cost drops to weeks, a security team can get a real proof on the parts that matter most.

A mathematical proof that cryptographic code works as specified is the highest bar in software security. Teams of experts have spent years building one for a protocol as small as Signal. Google DeepMind says AI agents did it for Signal in four weeks.

Google DeepMind introduced AlphaProof Code this week. It is an agentic framework that writes machine-checked proofs for software. In the first large test, agents powered by Gemini Carbon wrote more than six million lines of Lean. The proof covered key security properties of the actual code behind the Signal protocol. Signal protects messages on Signal, Google Messages and WhatsApp.

What the agents verified, and what they did not

DeepMind says the proof covers the compiled code that runs on devices. Most prior verification of cryptography stopped at the design layer. Bugs tend to live in the gap between a spec and the live code. DeepMind says this is the first time a widely used protocol's live code has been formally verified. That holds for any method, with or without AI.

Two security properties were proved. Messages stay protected before and after a device is briefly compromised. Lean checks every proof on every run. DeepMind's experts did not have to trust the agents' work.

The agents also ran a sanity check. The team put two real, long-fixed bugs back into an isolated copy of Signal's code. AlphaProof Code produced a machine-checked proof that the buggy version was insecure, in both cases.

Where the humans stayed

Human experts set the targets. They decided what to prove, wrote the formal spec with models, and reviewed every line. The agents then audited their own output during the proof. About one percent of the codebase was sent back for a second look by a person. Lean checked the rest on its own.

What it says for a team building software

DeepMind is open-sourcing the proofs and the Signal spec. Other teams can read exactly what was verified and copy the method. The technique is still slow and narrow. One protocol, in one language, with a spec written by cryptographers. A team running a normal web backend cannot try AlphaProof Code on its code tomorrow.

The direction matters anyway. An agent can now translate live code into a Lean theorem and prove it. The cost of mathematical assurance on the hardest parts of a system drops from years to weeks. That means the key exchange, the authentication layer, the audit log. For a team that already runs a serious security review, that is the part worth watching. Ehren Kret, Signal's chief technology officer, says the Signal Foundation now has higher confidence in its protocol.

The announcement reached 102 points on Hacker News.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.