AlphaProof Code checks proofs against human-reviewed specifications
Google DeepMind reports proving selected security properties of compiled libsignal code using AI agents and Lean. Humans review the mathematical specification; agents generate proofs that Lean checks against it. The work covers a specific AArch64 harness and assumptions about cryptographic routines, rather than proving the entire Signal application secure or removing the need for expert judgment.
Google DeepMind · Nov 10, 2026
Why it matters
Machine-checking generated proofs helps only when the specification and its assumptions express the guarantee people actually need.
More context
The public reviewer site links theorem statements to proof code. Security and verification engineers can inspect that mapping first. Rebuilding requires pinned Lean, Rust and supporting tools; the repository estimates several hours and roughly 15–34 GiB peak RAM. Linux is required for its comparator checks.
Keep in mind
Research artifacts, not a general-purpose security product. The trusted specification, machine semantics, Lean kernel and primitive contracts remain assumptions. This review covered the overview and specification sections, not all proof files or a reproduced verification run.
Original read