Back to LLM Models

DeepMind verifies Signal's real code with AI agents

Google DeepMind says its researchers guided AI agents to mathematically prove key security properties of the actual code underlying the Signal protocol, the encryption layer behind billions of private conversations in apps like Signal and Google Messages. The system is called AlphaProof Code, an agentic framework built for formal verification, meaning proving in mathematical terms that code behaves as specified. According to the company, the agents generated over six million lines of Lean, and the whole verification took four weeks.

Formal verification is the highest standard of security assurance, and it is difficult and labor-intensive. Experts have spent decades analyzing cryptographic protocols. Google DeepMind says this is the first time AI has been used to formally verify a widely deployed cryptographic protocol at this depth, and also the first time an implementation like Signal's has been formally verified by any method, with or without AI.

How the verification ran

AlphaProof Code splits the work the way a research team would. Human experts worked with a mix of models to translate Signal's security goals into formal mathematical specifications, and humans reviewed every one of them. From there, agents powered by Gemini Carbon did the proofs, generating the six million lines of Lean, the language used to machine-check mathematical claims about code.

Once the proofs were in, the agents audited them and flagged the parts that needed a second human look, a scope of about one percent of the codebase. Lean checks the rest automatically, so the experts' review concentrated on the parts where human judgment carries the weight.

Beyond the design documents

Most protocol verification stops at the design level: you prove the math of the protocol, not the code. Real-world vulnerabilities often hide in the gap between design and implementation. The announcement says AlphaProof Code proved that the compiled code fulfills the key security properties, including keeping messages protected before and after a device is temporarily compromised. The proof targets the exact code running on devices, which is what makes these proofs vastly more complex than the design-level kind.

The four-week figure is one the company reports for this task. Work at this scale has historically taken expert teams years.

Built to catch bugs, not just to pass

To confirm the system can tell correct code from buggy code, the researchers reintroduced two real, previously fixed bugs into an isolated copy of Signal's code. In both cases AlphaProof Code produced a machine-checked proof that the buggy version was insecure.

Ehren Kret, chief technology officer of Signal, framed the result as a matter of confidence. As AI-driven security threats grow, he says, having the core Signal protocol rigorously verified with automated theorem proving gives the team "even greater confidence that our code delivers on its privacy guarantees."

What to take from this

Google DeepMind is open-sourcing the formal verification artifacts for community use, alongside the technical paper. The stated next steps are broader systems, from medical device firmware to power grid software, and pairing formal verification with defense agents like CodeMender to match the speed and scale of agent-driven cyber threats.

For a team that runs production code handling sensitive data, the practical number is the cost curve. Formally verifying deployed cryptographic code used to be a multi-year research project. By the company's numbers it now fits in a quarter. The scope here is one protocol at one depth, and DeepMind itself frames scaling as continued research. The part that does not change: a machine-checked proof is the strongest assurance available for code, and the cost of getting one just came down.

Source: deepmind.google

More news

Back to LLM Models

Have something in mind?

Start a conversation