Article

Google has AI perform mathematical proofs for the Signal encryption protocol, completed in 4 weeks, and also found vulnerabilities in older versions

Dongcha Beating AI Flash News: Google DeepMind developed an AI system called AlphaProof Code, which used mathematical methods to verify some core security guarantees of the Signal encryption protocol in just 4 weeks.

Ordinary security testing can only check limited scenarios; not finding a vulnerability does not mean no vulnerability exists. Mathematical proofs are more rigorous: first write the security requirements that software must satisfy as mathematical formulas, then prove that these requirements always hold under specified conditions.

In the past, such proofs required experts to spend a lot of time deriving them. AlphaProof Code allows AI to attempt proofs on its own, modify when derivation issues are found, and continue verification. Human experts are responsible for reviewing key steps.

Researchers also tested it against older versions of Signal with known vulnerabilities. The AI successfully identified these already-fixed issues and confirmed that the older versions could not pass the corresponding mathematical proofs. For the current protocol parts checked this time, the AI did not find new vulnerabilities.