What was done
Google DeepMind presented AlphaProof Code, an artificial intelligence system designed to mathematically prove that software complies with its security specifications without errors. According to a report by The Decoder, researchers applied the system to Signal's encryption protocol, completing in four weeks work that normally takes years.
The central aspect, according to the blog of DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri cited by the source, is that the verification did not stop at the theoretical level of the design, as happened in previous attempts: AlphaProof Code analyzed the actual cryptographic code running on users' devices. After experts defined the formal security requirements, AI agents based on Gemini Carbon wrote more than six million lines in the Lean proof language, flagging about 1% of the codebase for human review.
Why it matters
The novelty concerns the shift from design to verification of real software: according to the source's reconstruction, this is where previous efforts had stopped. Formal verification, in general terms, aims to establish with mathematical rigor that the code complies with specifications, rather than limiting itself to empirical tests; the material does not clarify in detail the formal properties required or the results of the verifications beyond the two test bugs.
As a test of the method, the researchers deliberately inserted two bugs into the code: according to the source, they were identified immediately. The figure on human involvement — about 1% of the code flagged for review — is presented as an indicator of the degree of automation achieved, but the source does not specify how this percentage was measured or how many experts participated in defining the requirements.
Statements and outlook
Signal CTO Ehren Kret said, according to The Decoder, that the approach strengthens confidence in data privacy as AI-related threats grow. This is a statement reported by the source and not a result independently verifiable from the available material.
DeepMind intends to extend the method to critical infrastructure such as power grids and medical devices, through systems such as CodeMender. The timing, methods and limits of this extension remain to be clarified: the source does not provide details on applicability criteria or on any independent verification of the proof relating to Signal.
