Google DeepMind's AlphaProof Code proves Signal's security at the machine-code level
Google DeepMind's AlphaProof Code agents wrote about 6.5 million lines of Lean proofs that compiled libsignal keeps Signal's core security guarantees.
By Emre A.Published
Tech AI LatestGoogle DeepMind introduced AlphaProof Code on 10 November 2026, a multi-agent framework that writes formal specifications with humans and then proves code meets them in Lean. Applied to libsignal, the Signal protocol's reference implementation, its agents produced about 6.5 million lines of machine-checked proofs covering forward secrecy, authentication, post-compromise security and handshake secrecy for the compiled AArch64 binary. A team of five did the work in four weeks, and the proofs are open source.
Key facts
- Announced: 10 November 2026 (Google DeepMind), with a paper and Lean code on GitHub
- Target: unmodified libsignal v0.105.4 with SPQR v1.6.0, compiled with stock rustc to AArch64 Linux machine code
- Proofs: about 6.5 million lines of Lean, resting on about 26,000 lines of trusted specifications and AArch64 semantics
- Effort: a team of five over four weeks; no author wrote proof code, per the paper
- Models: specifications drafted with agents powered by Gemini 4 Carbon and Claude Opus 5.5; proofs by Gemini 4 Carbon agent teams
- Out of scope: group messaging, sealed sender, side channels; cryptographic primitives are assumed secure
What DeepMind built
Google DeepMind announced AlphaProof Code on 10 November 2026, together with a paper by researchers from Google DeepMind and Google and the full Lean development on GitHub. The framework has three parts: a human-guided agent that writes specifications in the Lean proof assistant, a formal semantics of AArch64 machine code, and a team of autonomous prover agents. A separate gatekeeper agent accepts a proof only if Lean's kernel checks it and the comparator tool confirms it proves exactly a held-out copy of the specification the humans approved.
The team applied it to libsignal, the official Rust implementation of the Signal protocol; variants of the protocol are also deployed in WhatsApp, Messenger and Google Messages, according to the paper. They linked libsignal v0.105.4 into a small command-line harness, compiled it unmodified with the standard Rust compiler and proved properties of the resulting machine code, about 140,000 AArch64 instructions. The proofs cover the post-quantum PQXDH handshake and the Triple Ratchet and establish forward secrecy, authentication, post-compromise security and handshake secrecy, in both the symbolic and the computational models of protocol security. DeepMind says these are the first such proofs for the unmodified code of a widely used cryptographic implementation.
The numbers and the method
According to the paper, the agents generated about 6.5 million lines of Lean proofs: 5.3 million about the binary, about 860,000 for the computational proof and about 350,000 for the symbolic one. What a reader has to trust is far smaller, about 18,000 lines of specifications plus about 8,000 lines of AArch64 semantics. Because the proofs are about the binary, the Rust compiler does not need to be trusted.
The work took a team of five four weeks. The paper says specifications were written in exchanges between the researchers and agents powered by Gemini 4 Carbon and Claude Opus 5.5, with a model from a different family reviewing each draft, while the proofs came entirely from Gemini 4 Carbon agent teams. For comparison, the authors cite the seL4 microkernel, which took twenty person-years to verify.
To test that the specifications catch real mistakes, the team reintroduced two previously fixed libsignal bugs and proved in Lean that the resulting binaries violate the symbolic security theorem. Writing the specifications also exposed errors in published analyses: the paper lists fifteen gaps in a computational analysis of the Double Ratchet that had to be repaired before the proofs went through.
Limits and what comes next
The guarantees are narrower than the headline. The proofs assume the cryptographic primitives are secure, trust the contracts of some hooked routines, cover one-to-one sessions only, and leave out group messaging, sealed sender and server-side protocols. The computational theorem is classical, so it does not prove the post-quantum healing SPQR is designed for. The authors say most of their effort went into specifications rather than proofs, and that specifications for widely deployed systems will need open review by the communities that rely on them.
The result adds to a run of AI-assisted proofs this autumn, from Anthropic's Lean work on Fermat's Last Theorem to Microsoft Research's KLS proof. DeepMind says it wants to verify new libsignal releases before they ship.
What changed
| Before | Formal security proofs of Signal covered abstract protocol models or modified reimplementations, not the shipped code |
|---|---|
| After | AI agents produced machine-checked proofs that the compiled, unmodified libsignal binary keeps key security properties |
| Who it affects | Signal and apps built on its protocol, security auditors, and teams weighing formal verification for critical code |
Questions people ask
- What is AlphaProof Code?
- A Google DeepMind framework, announced on 10 November 2026, in which humans and AI agents write formal specifications in Lean and autonomous agents then prove that compiled code meets them.
- Did AlphaProof Code find bugs in Signal?
- The paper reports proofs that the verified libsignal build satisfies the stated properties. To test the specifications, the team reinserted two old, fixed bugs and proved the modified builds insecure.
- Can anyone check the proofs?
- Yes. Google DeepMind published the Lean specifications and proofs on GitHub under the Apache 2.0 license; checking all of them takes several hours and about 15 to 34 GiB of RAM, per the repository.
Sources
- Introducing AlphaProof Code: a breakthrough in verifying complex software at depth (Google DeepMind, 10 Nov 2026)
- AlphaProof Code: Verifying Signal's Security at the Machine-Code Level (paper and Lean development) (Google DeepMind, 10 Nov 2026)
- google-deepmind/code-verification (GitHub, 10 Nov 2026)
- Google DeepMind melds humans and AI agents to vet Signal's secure messaging techniques (The Register, 10 Nov 2026)
Reference pages
- Gemini 4 Carbon model
- Claude Opus 5.5 model
- Google DeepMind company
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.





