AlphaProof Code verifies Signal’s cryptographic implementation
Google DeepMind says its agentic framework formally verified key security properties in Signal’s deployed code and released the proof artifacts.
What happened
Google DeepMind introduced AlphaProof Code, an agentic framework that formally verified key security properties in the code underlying the Signal messaging protocol. The work covers production cryptographic implementation rather than only an abstract protocol specification, and the researchers describe it as the first verification of a widely deployed protocol at this depth.
The system combines human-designed specifications with multiple AI agents powered by Gemini Carbon. Researchers translated Signal’s security goals into Lean, then used agents to generate more than six million lines of proof code. Lean, rather than the language model, served as the final checker: every accepted result had to satisfy machine-verified mathematical constraints.
The team also inserted two previously fixed vulnerabilities into an isolated version of Signal’s code. AlphaProof Code identified both as insecure, providing a test that the framework could distinguish a valid implementation from one that merely looked plausible.
Why it matters
Formal verification has long offered the strongest available assurance for security-critical software, but its cost and specialist demands have limited its use to relatively small components. AlphaProof Code suggests that agentic systems may expand the practical scope of theorem proving into large, actively deployed codebases. That matters well beyond messaging: the same workflow could eventually be applied to medical devices, industrial control systems and infrastructure software.
The claim remains bounded. Human experts still chose and reviewed the specifications, and the published result concerns a defined portion of Signal’s implementation rather than every possible property of the entire application. Even so, the combination of open-sourced artifacts and a real security protocol makes this more consequential than a benchmark demonstration.
Uncle Cat take
Six million Lean lines are impressive, but the strategic prize is specification review: AI can accelerate proofs only after humans decide what “secure” actually means.