AI agents prove Signal's core encryption code correct in four weeks

DeepMind says AlphaProof Code formally verified the real Signal protocol code, a job that has historically taken expert teams years.

Google DeepMind announced AlphaProof Code on November 10, 2026. It is a multi-LLM agentic framework for formal verification that pairs human experts with AI agents. DeepMind researchers used it to mathematically prove key security properties of the actual code underlying the Signal protocol, which protects private conversations in apps such as Signal and Google Messages.

Human experts reviewed every formal specification of Signal's security goals, which were drafted with a mix of models. Agents powered by Gemini Carbon then produced the proofs, generating over six million lines of Lean. The agents audited their own work and flagged about one percent of the codebase for human re-review, while Lean automatically checks the rest. The proofs cover compiled code and properties such as message protection before and after a temporary device compromise. In a test, researchers reintroduced two real, previously fixed bugs into an isolated version of Signal's code, and the system produced machine-checked proofs that the buggy versions were insecure.

Why it matters: formal verification gives stronger assurance than testing, but it has been slow and expensive. DeepMind says this run took four weeks, against years for expert teams historically. If that cost holds up, provably correct security software becomes something far more projects can afford, which widens access to trustworthy private communication. DeepMind says this is the first time AI has formally verified a widely deployed cryptographic protocol at this depth, and the first time such a protocol's implementation has been verified by any method. Signal CTO Ehren Kret called the use of AI agents to translate live code into Lean theorems a major breakthrough. DeepMind is open-sourcing its verification artifacts.

Caveats: this is DeepMind describing its own work, and it has not been independently verified. The verification covers Signal's core cryptographic code, not all of Signal. DeepMind suggests the approach could extend to medical device firmware and power grid software, but that is a stated possibility, not a demonstration, and it says scaling requires continued research.

Source: Google DeepMind