Google has AI create a mathematical proof for Signal's encryption protocol in 4 weeks, and also discovered vulnerabilities in older versions.

Beating AI News Flash: Google DeepMind has developed an AI system called AlphaProof Code that, in just four weeks, mathematically verified part of the core security guarantees of the Signal encryption protocol.

Ordinary security testing can only check limited cases; the absence of discovered vulnerabilities does not mean vulnerabilities do not exist. Mathematical proof is stricter: first, the security requirements that software must satisfy are written as mathematical formulas, and then it is proven that these requirements always hold under specified conditions.

In the past, such proofs required experts to spend a great deal of time deriving them. AlphaProof Code can let AI attempt proofs on its own, revise them when problems in the derivation are found, and then continue verification. Human experts are responsible for reviewing key steps.

Researchers also tested it against an old version of Signal with known vulnerabilities. The AI successfully identified these issues that had long since been fixed and confirmed that the old version could not pass the corresponding mathematical proof. For the parts of the current protocol examined this time, the AI found no new vulnerabilities.