Google DeepMind on November 10 announced AlphaProof Code — an AI agent framework that formally verified the cryptographic core of the Signal protocol, down to the compiled ARM machine code. That last detail is the news. Formal verification of Signal’s protocol design has existed for years. DeepMind’s agents proved the running binary.
The Gap Nobody Was Closing
Formal verification tools like ProVerif and Tamarin have long been able to prove properties of cryptographic protocol models — the mathematical blueprints. But a blueprint is not a building. Compiler behavior, memory layout, and ABI differences mean that a provably correct design can compile into a binary with new failure modes. This is not a theoretical concern: Heartbleed and a long line of TLS vulnerabilities trace back to the gap between protocol design and running implementation.
AlphaProof Code covered that gap. It verified libsignal v0.105.4 — Signal’s reference cryptographic library — compiled to AArch64 machine code. The same binary that runs on devices. DeepMind describes this as a historic first: the first time any method has formally verified a widely deployed protocol’s actual running implementation, not just its design.
What Was Proved
The verification covered two of Signal’s protocols. PQXDH (Post-Quantum Extended Diffie-Hellman) is the initial handshake used when two parties first contact each other; it blends classical Diffie-Hellman with a post-quantum key encapsulation mechanism to stay safe against future quantum computers. The Triple Ratchet is Signal’s newer ongoing key-rotation scheme, combining the classical Double Ratchet with the post-quantum SPQR for hybrid forward secrecy.
For both, AlphaProof Code established four properties: forward secrecy, authentication, post-compromise security, and handshake secrecy. As a stress test, DeepMind reintroduced two previously-fixed Signal bugs. AlphaProof Code caught both and showed the protocol could not be verified with those issues in place.
The Scale of the Work
The project took four weeks. Expert verification teams typically spend years on scope like this. AlphaProof Code’s agents — powered by Gemini 4 Carbon — generated roughly 6.5 million lines of Lean proof code, resting on approximately 26,000 lines of trusted specifications and AArch64 semantics. Lean’s kernel then checked every step. About 1% of the codebase was flagged for human review; the rest was verified automatically.
The architecture splits work the way a research team would: human experts write the security specifications (what to prove), and AI agents do the proving. The agents cannot just assert correctness — Lean rejects any proof that does not check out mechanically.
The Specification Is Now the Trust Boundary
This is the nuance that matters. AlphaProof Code proved that libsignal’s implementation satisfies its formal specification. It did not prove that the specification captures every security requirement worth caring about. If the spec is wrong or incomplete, the proof is still valid — and security may not be. The Hacker News thread on this release zeroed in on exactly that point within hours.
The team also published what was not covered: payload privacy, responder authentication for a first message in an incomplete handshake, one-time Kyber prekeys, the keygen operation itself, side channels, and group chats. That list is not a failure — no verification effort covers everything — but it is worth reading before concluding that Signal’s implementation is now fully proved. Qodo’s breakdown of the limitations is worth bookmarking.
What Developers Can Access Today
AlphaProof Code itself is not available as an API or product. It remains a research project. What DeepMind has open-sourced under Apache 2.0 is the full set of verification artifacts: the Lean specifications, the AArch64 formal semantics, and the theorem files. If you want to audit the proofs, run them in Lean, or build on the AArch64 semantics model, the materials are there. If you want to run AlphaProof Code on your own codebase, there is no mechanism for that yet.
The broader signal is harder to miss. Signal’s PQXDH specification and the full verification scope are publicly documented. And Anthropic’s formal verification research made the same point earlier this year: LLMs are making formal verification tractable at scales previously out of reach for all but the best-funded research groups. AlphaProof Code is the first production-scale demonstration of that claim. The question for the industry is how long before this becomes a workflow tool rather than a research milestone.

