AlphaProof Code: How DeepMind's AI Proved Signal Is Secure
AlphaProof Code is Google DeepMind's AI agent system that formally verified Signal's libsignal down to machine code in 4 weeks. How it works, what it proved and why it matters.
Quick answers
- What is AlphaProof Code?
- AlphaProof Code is a Google DeepMind agent framework for formal software verification, announced November 10, 2026. Humans and AI write the spec; Gemini 4 Carbon agents write the proofs; Lean checks every step.
- What did AlphaProof Code prove?
- It proved key security properties of libsignal, the Signal protocol's reference code, down to the AArch64 machine code: forward secrecy, authentication, post-compromise security and handshake secrecy.
- How long did AlphaProof Code take?
- Four weeks with a team of five. For comparison, verifying the seL4 microkernel by hand took about twenty person-years.
- Can I use AlphaProof Code?
- Not as a product yet. DeepMind open-sourced the Signal verification artifacts and published a technical paper.
AlphaProof Code is DeepMind's AI that mathematically proves real code is secure, and its first target was Signal. Google DeepMind introduced AlphaProof Code on November 10, 2026. Its agents proved that the compiled code of libsignal, the library that protects chats in Signal and Google Messages, keeps its core security promises.
DeepMind says this is the first time AI has formally verified a widely deployed cryptographic protocol at this depth, and the first time such an implementation has been verified by any method. Here is how AlphaProof Code works, what it proved and what it means.
AlphaProof Code at a glance
Item | AlphaProof Code |
|---|---|
Maker | Google DeepMind |
Announced | November 10, 2026 |
Type | Multi-agent framework for formal specification and verification |
Target | libsignal (Signal protocol reference code) |
Depth | AArch64 machine code, not just the design |
Proof size | About 6.5 million lines of Lean, ~140,000 instructions |
Trusted base | ~18,000 lines of specs + ~8,000 lines of AArch64 semantics |
Effort | 4 weeks, 5 people, no human-written proof code |
Prover | Agents powered by Gemini 4 Carbon |
How AlphaProof Code works
- Spec: researchers work with a mix of models to turn Signal's security goals into Lean specifications. A human expert reviews every one.
- Machine-code semantics: a formal model of AArch64 lets the proofs talk about the exact bytes that run on phones.
- Prover agents: a team of Gemini 4 Carbon agents writes the proofs. A proof is accepted only if Lean's kernel checks it.
- Audit: agents flag the parts that need a second human look, about 1% of the codebase.
AlphaProof Code builds on AlphaProof, the 2024 system that reached silver-medal level at the International Mathematical Olympiad by searching for proofs in Lean. The new framework points the same idea at software instead of math problems.
What AlphaProof Code proved about Signal
The proofs cover Signal's post-quantum PQXDH handshake and its Triple Ratchet. They show forward secrecy, authentication, post-compromise security (messages stay safe after a device is briefly hacked) and handshake secrecy hold in the compiled code, MLQ.ai summarizes.
To test that the system can tell good code from bad, DeepMind put two real, already-fixed bugs back into an isolated copy of the code. Both times, AlphaProof Code produced a machine-checked proof that the buggy version was insecure. Signal CTO Ehren Kret called using AI agents to turn live code into valid Lean theorems "a major breakthrough".
Why AlphaProof Code matters
- Speed: four weeks against about twenty person-years for the hand-verified seL4 kernel.
- The gap between design and code: most protocol proofs stop at the design. Real bugs often live in the code, which is what was proved here.
- Defense against AI attacks: DeepMind pitches it next to CodeMender as a way to keep up with agent-driven cyber threats, and names medical firmware and power-grid software as next targets.
- Gemini 4 Carbon in action: it is one of the first public jobs for Google's new frontier model. See our Gemini 4 Carbon guide.
Limits are real too. AlphaProof Code is a research system, not a product you can sign up for; specs still need expert humans; and the trusted base of about 26,000 lines still has to be right.
Put AI agents on your team's long jobs
AlphaProof Code shows what agent teams can do when the work can be checked. A Sumus AI Employee brings the same idea to everyday work: it takes a long task, splits it, checks its own output against your rules and reports back in your team chat.
FAQ
When was AlphaProof Code announced?
November 10, 2026, by Google DeepMind.
Which model powers AlphaProof Code?
The prover agents run on Gemini 4 Carbon.
Is AlphaProof Code open source?
DeepMind open-sourced the formal verification artifacts for Signal and published the technical paper; the framework itself is not a public product.
How is AlphaProof Code different from AlphaProof?
AlphaProof proves math problems in Lean; AlphaProof Code proves that real software meets a security spec.
Can a Sumus AI Employee help with long engineering tasks?
Yes. It splits work into steps, checks results and reports to your team chat.
Update log
- — Nov 11: First published. AlphaProof Code announced Nov 10, 2026; Signal libsignal verification details.