Audit notes
We checked DeepMind’s AlphaProof Code release
Published · Updated
We checked DeepMind’s AlphaProof Code against the public repo google-deepmind/code-verification at commit 808b4b7ffef198e6a91f71ed776d641a1c236882.
The release holds 11,323 Lean files and 7,528,293 raw Lean lines — about 7.5 million, which fits “over six million lines,” not the “three million” going around. We read the released files. Here’s what the proofs actually cover.
This note is a static scan only. We have not run the kernel build, we make no kernel-checked claim, and we report no planted-bug results.
What the release claims
The DeepMind blog says the proofs reach “down to the exact code running on devices.” The paper is more careful: what’s verified is their AArch64 Linux libsignal-cli build, “not Signal’s own applications.” The README describes “two AI-generated proofs of security for compiled AArch64 machine code of libsignal … in a CLI harness (with hooked cryptographic primitives).”
Android and iOS are covered only “as long as we trust” those builds preserve the behavior of the protocol code as exercised by the libsignal-cli binary.
What is hooked
37 routines are hooked: swapped for contracts instead of run. 23 are crypto primitives (SHA-2, AES-CBC, HMAC, X25519, Curve25519 point and scalar ops, Kyber1024, ML-KEM-768), which the paper says it idealizes. That’s standard practice, and it means the result is conditional on those primitives being secure.
The other 14 hooks aren’t crypto: libc memmove, bcmp, and strlen; the two output writes; Rust-std leftovers including memrchr, UTF-8 conversion, and error-message formatting; a fixed clock; an optimization barrier; and zeroizing secrets on drop. The paper says these contracts “are not proved against the real routines,” though DeepMind checked that its test scenarios give the same outcomes as on the real binary. Key generation is out of scope.
| Hook class | Count | What that means |
|---|---|---|
| Cryptographic primitives | 23 | Idealized; results are conditional on their security |
| Non-crypto contracts | 14 | “Not proved against the real routines” |
| Total hooked | 37 | Replaced by contracts at the binary boundary |
What we checked
Ford’s light receipts pass (America/Denver, Nov 10, 2026) pinned the repo at 808b4b7…882 and ran a Lean-aware masked scan (comments/strings masked) over the tracked .lean corpus.
| Gate | Result |
|---|---|
| Pinned corpus | 11,323 .lean files; 7,528,293 physical newlines; toolchain leanprover/lean4:v4.34.0; Rust 1.98.1; libsignal v0.105.4 / SPQR v1.6.0 |
| Hook table | 37 hooks in cliHookTablePlaced (23 cryptographic, 14 non-crypto) |
sorry in headline theorem folders |
None in Computational/Theorems/, Symbolic/Theorems/, or LibsignalBinary/ |
TRUSTED_BASE.md |
Missing — cited as the trust-banner target, absent from the tree at this commit |
| Key generation | Out of scope; theorems cover init / send / recv and quantify over provisioned keys |
lake build / #print axioms |
NOT RUN — static pass only |
The scan did find 43 sorry hits in 22 files, all outside those headline folders (testing fixtures, challenge stubs, and other non-headline trees). A static scan does not show that the headline theorems are sorry-free under a kernel check.
What we are and are not saying
We are saying the released artifact is large and checkable, that 11,323 Lean files and about 7.5 million raw lines fit “over six million” rather than “three million,” that 37 routines are hooked (23 idealized crypto primitives and 14 non-crypto contracts “not proved against the real routines”), that a static scan at 808b4b7 found no sorry in the headline theorem folders, that key generation is out of scope, and that TRUSTED_BASE.md is missing from the repo.
A fair summary is a Lean security proof for a compiled CLI build of libsignal with modeled crypto (we haven’t rebuilt it ourselves yet). That’s a real step, and it’s not “Signal on your phone.”
We are not saying we kernel-checked the proofs, that the contracts match the real routines, that the result applies to Signal’s Android or iOS apps without the extra trust assumption, or that we reproduced any planted-bug result. Line counts describe the corpus we scanned; they are not a measure of verified mathematics.
Sources and pinned receipt
- DeepMind blog: https://deepmind.google/blog/introducing-alphaproof-code/
- Repo: https://github.com/google-deepmind/code-verification @
808b4b7ffef198e6a91f71ed776d641a1c236882 - Paper / reviewers’ site: https://google-deepmind.github.io/code-verification/
- Toolchain:
leanprover/lean4:v4.34.0; Rust1.98.1; libsignal v0.105.4 / SPQR v1.6.0
Receipts: Ford’s RECEIPT.md, hooks_extracted.json, sorry_hits.json, and scan_out.json for the C18 AlphaProof Code static pass (Nov 10, 2026, America/Denver). Verdict: static scan only; no sorry in headline theorem folders; kernel build not run; no planted-bug results. Ford interest: 8/10.