A proof about compiled code is useful only if you know what behavior it promises.
On November 10, Google DeepMind introduced AlphaProof Code, a system that uses agents powered by Gemini 4 Carbon to produce formal proofs. The researchers applied it to selected compiled AArch64 code from libsignal, the library behind the Signal protocol. DeepMind says the work took about four weeks.
What Lean checks
The technical report describes three separate pieces: the specification, a model of the compiled code, and a proof linking them. The specification states the behavior to check. The agents propose proofs, and Lean checks whether each proof follows from the definitions and assumptions. People review what the specifications actually say.
Checking compiled code lets researchers reason about the instructions that run in the selected build. The result still depends on the model of those instructions. It also depends on how the proof treats calls to external routines.
Read one assumption before trusting the result
Pick one result in the published verification repository. Find the entry point and the conditions the proof assumes. Then look at how it describes an external routine. If that routine is supplied as an assumption, check whether another proof covers it or whether it sits outside the result.
The same process helps a team plan a pilot. Start with one function that has a clear rule, such as rejecting input of the wrong length. Record the build and architecture. Ask the reviewer to read the rule before looking at the proof. DeepMind’s choice of model is one part of the project; see our Gemini 4 Carbon coverage for its separate API rollout.
The results apply to the selected code and scenarios. They do not certify the security of the Signal app or cover every client, platform, or possible execution.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
