DeepMind's AlphaProof Code verifies Signal in four weeks

Proving code correct used to cost years. Google DeepMind says its agents just did it for Signal in four weeks.

In a post dated Nov. 10, Google DeepMind introduced AlphaProof Code, a system of AI agents that writes formal proofs about real software, and showed it on the code behind the Signal protocol.

Six million lines of proof, checked by a machine

The target was the compiled code underlying the Signal protocol, which DeepMind says protects billions of private conversations in Signal and Google Messages.

The split of labor is the point. Human experts decide what to prove. Researchers used a mix of models to translate Signal's security goals into formal specifications, and human experts reviewed each one, per the post. Then agents powered by Gemini Carbon wrote the proofs, and Lean, a proof checker, verified every one.

The numbers DeepMind reports:

• Over six million lines of Lean.
• About 1% of the codebase flagged by the agents for a human's second look.
• Four weeks, for work DeepMind says has historically taken expert teams years.

The proofs cover key security properties, such as keeping messages protected before and after a device is temporarily compromised, down to the code running on devices.

DeepMind also ran a test. It put two real, previously fixed bugs back into an isolated copy of Signal's code. Both times, the system produced a machine-checked proof that the buggy version was insecure.

"Using AI agents to translate live code into valid Lean theorems is a major breakthrough," Signal CTO Ehren Kret said in the post. A technical paper describes the work.

The bottleneck was always people

Formal verification is the strongest guarantee software can get: a mathematical proof that code does what its specification says, for every input, not just the ones you tested. It has always been rare for one reason. Writing proofs takes specialists, and they are slow and few.

AlphaProof Code moves the specialists up a level. They write and review what must be true. The agents grind out the proof. A checker that can't be fooled by confident prose decides whether it holds.

That last part is why this one deserves your attention. An AI that claims your code is secure is worth little. An AI that hands you a proof a machine has checked is worth a lot.

What it means if you ship security-critical code

You can't download AlphaProof Code. DeepMind says it is open-sourcing its formal verification artifacts from the Signal work at its project page, but the post names no access program for the system itself.

So the near-term change is cost. If proofs get this cheap, "formally verified" stops being rare and starts showing up in vendor security reviews. Ask for it.

Note the limits. Humans still wrote and reviewed the specifications, and a proof only covers what the spec says. A wrong spec, proven perfectly, is still wrong.

The model behind the agents is the same Gemini Carbon family that Google put on sale to developers this week.

Watch for the second target

DeepMind names medical device firmware and power grid software as next, and says scaling these techniques requires continued research. It also proposes pairing verification with CodeMender, its AI code-security agent.

The proof that matters is the second system, verified by someone outside Google.

Four weeks, once. Watch for twice.