Google DeepMind’s AI Proves Signal’s Encryption With a 6.5-Million-Line Math Proof

By Fatima Raza | November 12, 2026

Google DeepMind has used artificial intelligence to formally prove that the encryption code behind Signal actually does what it promises, and the proof runs to about 6.5 million lines of machine-checked mathematics. The work, published on 10 November 2026, verified key security properties of the compiled code behind Signal’s encryption, not just the mathematical model on paper. No new vulnerabilities were found in the code examined, and Signal’s chief technology officer says the result gives the company even greater confidence in its privacy guarantees.

Key Facts

What DeepMind Proved

Cryptographers have studied Signal’s encryption for years, and the maths was already well understood. What DeepMind did is different: instead of checking the protocol as an abstract design, AlphaProof Code checked the actual compiled code that runs on a phone, libsignal version 0.105.4 with the SPQR v1.6.0 crates, compiled with stock Rust for ARM Linux in a small test harness.

Two parts were checked. The first is the PQXDH handshake, the post-quantum handshake that sets up a conversation. The second is the Triple Ratchet, which rotates the encryption key for every message. For one-to-one sessions, the proof establishes four properties: forward secrecy, so past messages stay unreadable even if a key is exposed later; authentication, so you are really talking to the right person; post-compromise security, so a conversation can heal after a key is compromised; and handshake secrecy, so setting up the conversation leaks nothing to an eavesdropper. DeepMind calls it the first AI-assisted formal verification of a widely deployed cryptographic protocol at this depth, and the first implementation-level verification of such a protocol by any method. Signal CTO Ehren Kret said it “gives us even greater confidence in our code’s privacy guarantees.”

How the Proof Was Built

The proof totals about 6.5 million lines of Lean, a language in which every step of a mathematical argument is checked by a computer, leaving no room for a skipped case. It was generated in about four weeks by teams of 40 to 80 AI agents powered by Gemini Carbon, running in stretches of 24 to 72 hours. Humans stayed in the loop: people wrote the formal specification defining what “secure” means, reviewed about 1 percent of the material, and kept final approval.

What has to be trusted is far smaller than the proof itself: about 26,000 lines of specifications, plus a model of how an ARM processor runs code. Everything else is checked by machine. DeepMind also tested its own system: researchers reintroduced two previously fixed bugs, “weak state” and “small subgroup”, into an isolated copy, and AlphaProof Code correctly proved the buggy builds insecure.

What It Does NOT Prove: The Limits

The verification covers one-to-one sessions only; group messaging is future work. Only the libsignal build was verified, not the full Signal app, the servers, the key servers, the user interface, the safety numbers, or sealed sender. Key generation was not verified, the post-quantum healing of SPQR was not proved, and the low-level cryptographic primitives underneath it all were assumed secure rather than proven secure.

None of this makes the result unimportant. The proof covers the heart of the protocol, the code that sets up conversations and rotates keys, while the edges still rely on other testing and review. The honest version is narrower and still remarkable.

Why It Matters for WhatsApp Users in Pakistan

Signal itself has a modest user base in Pakistan, but the protocol it pioneered is everywhere. The same Signal protocol protects chats in WhatsApp, in Messenger and in Google Messages, and WhatsApp is Pakistan’s dominant messenger, used by families, businesses and journalists alike. When the core code of that protocol is formally verified, the benefit reaches far beyond Signal’s own users. New features keep arriving, as our guide to WhatsApp message scheduling shows, and a verified protocol is the quiet foundation underneath.

The timing fits a wider pattern: AI is moving from writing code to checking code. Apple is building AI deeper into the home, as covered in our report on Apple HomeView’s launch, price and preorder details, and Android keeps gaining AI features, as in our ColorOS 17 launch coverage. DeepMind’s proof is the same trend applied to security: AI that proves software correct, not just generates it.

What Happens Next

DeepMind has released the specification and the proofs as open source, so independent researchers can inspect and challenge them. A private proof nobody else can check would be worth very little. The AlphaProof Code framework itself, however, has not been released as a general package, so other teams cannot download it and run it over their own code today.

The obvious next steps are also the hardest. Group messaging is the big gap, and extending the proof there would matter enormously for WhatsApp and Signal users alike, as would verifying key generation and the post-quantum healing left out this time. But a few weeks of AI-assisted work produced a proof at a scale unthinkable by hand, for the protocol that protects the private conversations of billions of people.

Frequently Asked Questions