Claimed AI-led

Lean proofs of security properties for the compiled machine code of libsignal

Google DeepMind's AlphaProof Code, a framework in which humans vet Lean specifications and Gemini-powered agents write the proofs, produced about 6.5 million lines of Lean establishing forward secrecy, authentication, post-compromise security and handshake secrecy for Signal's PQXDH handshake and Triple Ratchet as compiled to AArch64 machine code from unmodified libsignal, along with machine-checked disproofs for two previously fixed bugs reintroduced into the code.

Model
AlphaProof Code agents powered by Gemini Carbon
Field
Computer science
Date
2026-11-10
Human collaborators
Swarat Chaudhuri, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger, Karn Seth, Jonathan Katz, Charles Sutton, Pushmeet Kohli

What was found

The verified program is libsignal-cli, a small harness around the unmodified libsignal v0.105.4 and SPQR crates built by stock rustc, and the theorems are stated over whole runs of its AArch64 executable, so neither the Rust source nor the compiler is trusted. Humans and models wrote the specifications together, with a human reviewing each one; prover agents then wrote proofs that a gatekeeper accepted only when Lean's kernel checked them against the approved statement using standard axioms, with sorry and native_decide rejected. Reviewer agents flagged about one percent of the proof code for human review, and the work took four weeks. The two reintroduced bugs, a small-subgroup key encoding and a weak-state splicing attack, were found earlier by Cheval, Jacomme and Richards; for each the system proved the buggy binary insecure. The Lean development is public.

Novelty check

arXiv API, 2026-11-12: abs:libsignal OR ti:"Signal protocol" returned 11 records; the nearest is a September 2026 runtime-monitoring paper (SpecMon) that checks observed runs of Signal Desktop and WhatsApp Web against Tamarin models, which does not prove properties of the code. The paper's related work names the prior closest efforts: LibSignal* in F*, a separate implementation written for verification; Arquint et al., an end-to-end verified WireGuard key exchange derived from modified official code; hax panic-freedom proofs for the SPQR crate; and design-level ProVerif, CryptoVerif and Tamarin analyses of Signal and PQXDH. No earlier entry in this registry covers protocol verification of deployed code.

Caveats and known objections

Entered at claimed as a lab announcement, although the Lean development is public; nobody outside Google has yet rebuilt it or audited what it states. The guarantee is only as good as its trusted base, about 26,000 lines of specifications and AArch64 semantics, plus the contracts assumed for the cryptographic primitives and a few library routines, the Lean model of how Linux runs a process, and the tooling that extracts symbol names from object files. Key generation is not verified, side channels and secrets outside the modelled leaks are out of scope, and the forward-secrecy and post-compromise theorems do not rule out a known replay-based loss of forward secrecy. No new vulnerability was found: the two bugs were known, fixed and deliberately reintroduced. Signal's chief technology officer is quoted in support and Signal staff commented on the method, which is endorsement rather than an outside check. Autonomy is ai-led rather than autonomous because humans decided what to prove and reviewed every specification; the proofs themselves were written by the agents.

No independent check is recorded on this entry yet.

Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.

Or on GitHub: submit a check challenge the grade send a correction or send a pull request

Entry history (1 event)
  1. AddedEntered the registry graded Claimed and AI-led.

Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.

Community discussion

Graded claimed for verification and ai-led for autonomy. What these mean.

Cite this entry

Plain text
whataifound.org. (2026). Lean proofs of security properties for the compiled machine code of libsignal. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-11-10-alphaproof-code-libsignal
BibTeX
@misc{whataifound-googledeepmind-2026-libsignal,
  title        = {Lean proofs of security properties for the compiled machine code of libsignal},
  author       = {{whataifound.org}},
  year         = {2026},
  howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
  note         = {Result by Google DeepMind. Verification: Claimed. Autonomy: AI-led.},
  url          = {https://whataifound.org/finding/2026-11-10-alphaproof-code-libsignal}
}

Related findings

← All computer science findings in the registry