AI Trend Research
EN

← November 10, 2026 briefing · added in the 00:00 KST update

P1 Research Paper AI securityAgentsScience & bio00:00 updateGoogleSignal

Google DeepMind unveils AlphaProof Code; AI agents formally verify Signal's core cryptographic code down to machine code with Lean proofs

Announced November 10, 2026

Verified (fact)

Google DeepMind unveiled AlphaProof Code. It said the system's agents wrote 3 million lines of proofs in Lean to verify that Signal's core cryptography works correctly. According to a GeekNews summary, the verification target was libsignal, the reference implementation of the Signal Protocol. It was compiled to AArch64 Linux machine code and then verified. The verified security properties include two: messages exchanged before a temporary device compromise remain secure, and protection is restored after a compromise. The proofs were written by Gemini Carbon agents, starting from mathematical specifications that had been reviewed by experts. Lean automatically checked the correctness of the proofs. The verification reportedly took four weeks. The proof size differs between sources: 3 million lines in the official summary versus more than 6 million lines in the GeekNews summary.

Needs fact-check (claim)

Google DeepMind described it as a breakthrough in verifying complex software at a deep level.

Why it matters (interpretation)

It is a case of formally verifying a widely used, real-world cryptographic library at the level of compiled machine code rather than source code. If AI agents can produce large-scale formal proofs, the cost of verifying security-critical code could drop sharply.

Sources

More about Google