Google DeepMind AI formally verifies Signal protocol code in 4 weeks

Summary

• Google DeepMind says AI agents mathematically proved key security properties of Signal's underlying protocol code.
• Formal verification of a deployed cryptographic protocol's implementation had never been achieved before, by any method.
• The work compressed into four weeks a task the company says historically consumed expert teams years.

The latest

AI agents have mathematically proved key security properties of the actual code behind the Signal protocol, the encryption layer used in Signal and Google Messages, Google DeepMind said in a blog post published November 10, 2026. The company called it the first time any method, with or without AI, has formally verified the implementation of a widely deployed cryptographic protocol at this depth.

Details

• The tool: The work was done with AlphaProof Code, described by Google DeepMind as a multi-LLM agentic framework for formal verification. It splits the job the way a research team would: human experts decide what to prove, and a group of AI agents produces the proofs. The post was authored by Pushmeet Kohli and Swarat Chaudhuri.
• The scale: Agents powered by Gemini Carbon generated more than 6 million lines of Lean, the language used to mathematically prove code correctness. Google DeepMind did not detail the compute used or the number of agent runs behind that output.
• The human role: Researchers worked with a mix of models to translate Signal's security goals into formal mathematical specifications, with human experts reviewing every specification. The agents then audited their own proof and flagged the sections needing a second human look.
• The 1% figure: According to Google DeepMind, only one percent of the codebase was routed back for human review, with Lean automatically checking the remainder. The company framed that as a substantial time saving, letting experts concentrate where human judgment matters most.
• The timeline: The full verification process took four weeks, compared with the years Google DeepMind says expert teams have historically needed for comparable work. No start date for the project was given in the announcement.
• Design versus code: Most protocol verification stops at the design level, according to the company. The agents went further, proving that the compiled code itself fulfills the key security properties, which is the distinction the company says makes the result a first.
• Why protocols: Cryptographic protocols underpin nearly all digital communication, from online banking to web browsing, Google DeepMind noted. A single flaw can expose millions of users at once, which is why experts have spent decades formally analyzing protocol designs.
• The claim's limits: Google DeepMind described the result as covering key security properties rather than the entire protocol. It did not say whether Signal itself participated, whether the Lean proofs will be published, or whether the framework will be released beyond its research teams.

Background

Formal verification means mathematically proving software runs without flaws, the highest standard of security assurance available. It is labor-intensive, which is why it has mostly been applied to protocol designs on paper rather than to the code that billions of devices actually run.

Between the lines

The headline number is the one percent, not the six million lines. Google DeepMind's framing is that Lean's automatic checking turns AI output into something a small human team can supervise — the bottleneck moves from writing proofs to choosing what to prove and auditing the exceptions. That is the economic argument for applying the framework to banking or browsing protocols next, the uses the company itself cited.

What's next

Watch for publication of the Lean proofs and formal specifications, any response from Signal, and whether Google DeepMind applies AlphaProof Code to a second deployed protocol or opens the framework to outside researchers.

Source: Google DeepMind Blog