AlphaProof Code verified Signal cryptography down to machine code

Google DeepMind introduced AlphaProof Code, a multi-agent framework for formal software verification. Agents powered by Gemini Carbon produced more than six million lines of Lean proofs and, in four weeks, verified selected security properties of Signal’s core cryptography down to code running on devices.

The proof reached from design to compiled code

Human experts first translated security goals into formal specifications and approved each one. Agents then generated proofs checked mechanically by Lean. An audit flagged about 1% of the code for human review. The system also detected two previously fixed bugs that researchers reintroduced into an isolated version.

AI compresses the most expensive verification work

The value is not an automatic declaration that software is safe. Agents accelerate the translation of implementation into theorems and the search for proofs, while people decide what must be proven. That could bring formal methods within reach of teams unable to fund years of specialist work.

Mathematical certainty covers only specified properties

Lean checks every proof step, but it does not choose the right specification. A mistaken security goal or code outside the verified scope can remain uncovered. This is one team’s result on one demanding protocol, not a universal finished product.

Other repositories will determine practical value

DeepMind released the artifacts and technical paper. Independent reproduction and tests on firmware, libraries and other protocols will show whether the method travels without exceptional support from its authors. The effort required to maintain proofs after each code change will matter too.

Lilith's verdict

Lean checked millions of steps, but a person still drew the first line on the map. Draw that line wrongly and a perfectly precise route reaches the wrong place.