Google DeepMind uses AI to formally verify part of Signal’s encryption code

Executive Briefing
“Google DeepMind researchers used AI to mathematically verify part of the code implementing...”
Key Takeaways
- 1.The DeepMind team used AlphaProof Code, building on earlier verification work by the Signal Foundation, to examine part of Signal’s underlying code. It did not verify the app’s entire codebase.
- 2.People worked with AI agents to identify what needed to be proved. Multiple agents running Gemini Carbon then carried out the proof work and audited the proofs, flagging areas for human review.
- 3.The researchers said the verification took a month, compared with work that might ordinarily take experts years.
- 4.The team also tested whether AlphaProof Code could distinguish correct from flawed code using two previously fixed Signal bugs. The report says the tool produced correct proofs in that test.
Google DeepMind researchers used AI to mathematically verify part of the code implementing Signal’s encryption. The result adds confidence in the code they examined, but does not establish that the entire app is free of errors.
Key Developments and Technical Context
The DeepMind team used AlphaProof Code, building on earlier verification work by the Signal Foundation, to examine part of Signal’s underlying code. It did not verify the app’s entire codebase.
People worked with AI agents to identify what needed to be proved. Multiple agents running Gemini Carbon then carried out the proof work and audited the proofs, flagging areas for human review.
The researchers said the verification took a month, compared with work that might ordinarily take experts years.
The team also tested whether AlphaProof Code could distinguish correct from flawed code using two previously fixed Signal bugs. The report says the tool produced correct proofs in that test.
Strategic Analysis and Industry Implications
The notable step is AI’s role in both producing and auditing proofs, while people still define the verification goals and review flagged areas. If the approach transfers to other projects, it could reduce the work required to verify security-critical code.
WhatsApp and Google Messages using the same encryption does not mean their implementations were verified in this work. The reported result concerns the Signal code the team examined.
Outlook and Key Factors to Watch
The supplied account does not fully specify the technical boundaries of the verified code or which parts of Signal remain outside its scope.
It remains unclear how the process would perform on other systems. According to the report, DeepMind has not announced plans to turn the work into a consumer product.







