DeepMind's AI agents mathematically verify Signal's encryption code in four weeks

Google DeepMind has unveiled AlphaProof Code, a multi-agent AI framework for formal verification — mathematically proving that software does what its specification says. Its first big test was libsignal, the reference implementation of the Signal protocol that protects messages in Signal, Google Messages and other apps.

What was proved

• Human experts, working with AI models, wrote formal specifications of Signal’s security goals and reviewed each one.
• Prover agents (DeepMind says they were powered by Gemini Carbon) then generated about 6.5 million lines of Lean proof code, resting on roughly 26,000 lines of trusted specifications and a formal model of AArch64 machine code.
• The proofs cover the post-quantum PQXDH handshake and the Triple Ratchet, establishing forward secrecy, authentication, post-compromise security and handshake secrecy — for the compiled machine code of unmodified libsignal v0.105.4, not just the protocol design.
• The project took four weeks. As a check, the team reintroduced two real, long-fixed bugs into an isolated copy, and the system proved those versions insecure.

Limits

Cryptographic primitives and a few library routines were treated as trusted contracts rather than verified. The work did not cover the full Signal mobile app, and no new vulnerabilities were found. WIRED notes it is unclear how practical and affordable such verification will be at scale; Google did not disclose the cost. The Lean development has been published openly.