Skip to main content
The Quantum Dispatch
Back to Home
Cover illustration for AlphaProof Code: How AI Proved Signal's Encryption Code

AlphaProof Code: How AI Proved Signal's Encryption Code

Google DeepMind's AlphaProof Code generated about 6.5 million lines of Lean proofs to formally verify Signal's protocol code in four weeks, not years.

Kai Aegis
Kai Aegis★Nov 10, 2026★4 min read

Encryption only protects you if the code doing the encrypting matches the math on paper. On November 10, 2026, Google DeepMind announced AlphaProof Code, an agentic framework that formally verified key security properties of the actual code behind the Signal protocol, the encryption that protects billions of private conversations in apps like Signal and Google Messages. It is a genuine milestone for formal verification, and a strong example of AI being pointed squarely at defense.

  • What was verified: libsignal, the reference implementation of the Signal protocol, including its post-quantum PQXDH handshake and Triple Ratchet, per The Register
  • Scale: agents running on Gemini 4 Carbon generated roughly 6.5 million lines of Lean proofs, all machine-checked
  • Speed: five researchers finished the proof work in four weeks, a job Google says has historically taken expert teams years
  • Openness: Google is releasing the specifications and proofs publicly

What Is Formal Verification, in Plain Terms?

Most security testing looks for bugs. Formal verification goes further: it mathematically proves that software meets a written specification, so whole classes of flaws are ruled out rather than merely not found. It is considered the gold standard of security assurance, but it is notoriously expensive. The Register points to the seL4 microkernel, which took about 20 person-years to verify. That cost is why most cryptographic protocol analysis stops at the design level and never reaches the shipping code.

How Did AlphaProof Code Verify Signal?

According to Google DeepMind, AlphaProof Code splits the work the way a research team would. Human experts work with AI models to translate Signal's security goals into formal mathematical specifications, and humans review every one. SecurityWeek reports those specification agents ran on Gemini 4 Carbon and Claude Opus 5.5. Then Carbon-powered agents wrote the proofs in Lean, a language whose checker rejects any invalid step. Finally, the agents audited their own output and flagged roughly 1% of the work for a human second look.

The proven guarantees include forward secrecy, authentication and post-compromise security, meaning messages stay protected even before and after a device is temporarily compromised. Crucially, they hold for the compiled machine code, which closes the gap between design and implementation where real-world vulnerabilities often hide.

Can It Tell Good Code From Buggy Code?

Yes, and the team tested exactly that. They reintroduced two real, previously fixed libsignal bugs into an isolated build, and in both cases AlphaProof Code produced a machine-checked proof that the buggy version broke the security guarantees. That is the kind of result that turns a research demo into a credible engineering tool.

Signal's CTO, Ehren Kret, said in Google's announcement that rigorously verifying the protocol implementation gives the team even greater confidence in its privacy guarantees.

What Are the Limits?

The researchers are careful about scope, which builds trust. SecurityWeek notes the results cover Google's own build of the code run through a small command-line wrapper, not Signal's apps themselves, and side channels were excluded from parts of the analysis. The Register adds that the proof assumes perfect underlying cryptography. Those caveats are standard in formal methods and don't diminish the achievement.

Why This Matters for Defenders

DeepMind researcher Swarat Chaudhuri told The Register the dream is to scale this approach to code everywhere, with specification, not proof, as the next frontier. Paired with automated patching agents, provably correct software could become affordable for far more than a handful of projects. Signal itself has a strong track record here, as our look at Signal's automatic key verification showed. The model behind the work just opened to developers too; see our Gemini 4 Carbon launch coverage. For more defensive security news, visit our AI security section.

Sources: Google DeepMind: Introducing AlphaProof Code — November 10, 2026; The Register — November 10, 2026; SecurityWeek — November 10, 2026.

More Ai Security Stories