Google DeepMind says its AlphaProof Code system has formally verified the cryptographic core code used by Signal. The project took four weeks, applying mathematical proofs to software that runs on users’ devices rather than stopping at the protocol’s theoretical design.
From security requirements to code proofs
Formal verification starts with experts defining what a system must guarantee. For this project, researchers set security requirements for Signal’s encryption protocol, then AI agents built on Gemini Carbon worked to show that the implementation met those specifications.
The system produced more than six million lines in the Lean proof language. That volume reflects the detailed mathematical work involved in connecting software behavior to security requirements. It does not mean every line of the application was examined: the report says roughly one percent of the codebase was flagged for human review.
DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri described the work in a blog post. They said earlier verification efforts had reached the theoretical design stage, while AlphaProof Code checked the cryptographic core code running on users’ devices.
What the Signal project demonstrated
The project’s reported four-week timeline stands out because this kind of verification typically takes years. The result suggests AI agents can help with the labor-intensive process of producing formal proofs for real software, once people have specified the properties the software is expected to satisfy.
The researchers also tested the system against two bugs planted deliberately for the exercise. AlphaProof Code caught both immediately. That is evidence of how the system performed on those test cases; the source does not describe this as proof that the entire application is free of defects.
Human review remains part of the process. The flagged portion of the codebase gives experts areas to examine, while the formal proofs provide a way to assess whether the specified security properties hold in the verified code. Together, those steps offer a structured approach to checking security-sensitive software.
Why implementation-level verification matters
A protocol’s design can be sound while its implementation still contains mistakes. Verifying code on users’ devices addresses a different layer of the security question: whether the software that carries out the cryptographic work follows the requirements experts set.
That distinction matters for privacy tools. Signal’s encryption protocol is intended to protect communications, and checking its cryptographic core can strengthen confidence in how that protection is implemented. Signal CTO Ehren Kret said the approach strengthens trust in data privacy as AI threats grow.
Formal verification has a defined scope: it checks code against stated requirements. The quality and coverage of those requirements shape what the proofs can establish. The reported results therefore describe a specific verification effort on Signal’s cryptographic core, alongside human review of flagged code.
Where DeepMind wants to apply the method
DeepMind plans to extend the approach to other critical systems, including power grids and medical devices, through systems such as CodeMender. These settings also depend on software behaving as intended, though each would require its own security requirements and verification work.
The Signal project offers an example of AI-assisted formal verification applied to deployed cryptographic code. Its scale, timeline, and detection of the planted bugs show what the system accomplished in this case, while the human review process remains an important part of assessing the results.