Google DeepMind says AlphaProof Code helped prove selected security properties of the code underlying the Signal protocol, including properties of compiled code. The project covered specified parts of the protocol implementation; it did not formally verify the entire Signal mobile app, according to WIRED’s report. Google DeepMind’s announcement describes the method and the properties examined.
A formal proof applies to the code and security properties it actually covers.
What the proof covered
Formal verification turns a security goal into a precise mathematical statement, then checks a proof of that statement. Google DeepMind says its researchers proved selected properties of the Signal protocol’s actual code, including compiled code. One example was message protection before and after a device is temporarily compromised.
That scope matters: a proof about specified properties of protocol code is not a proof of every part of the Signal app. WIRED reported that the app as a whole was outside the analysis.
How people and AI agents worked together
Google DeepMind describes a process that began with researchers and models translating security goals into formal specifications. Human experts reviewed every specification. AI agents powered by Gemini Carbon then generated more than six million lines of Lean proof code. Lean is a proof assistant that checks whether a proposed proof follows from its formal rules.
The company says agents flagged 1% of the codebase for additional human review, while Lean checked the rest. That division describes this project’s reported workflow: people defined and reviewed the properties, agents generated proofs, and Lean checked them mechanically.
What the two-bug test showed
Google DeepMind says researchers reintroduced two previously fixed bugs into an isolated version of the Signal code. AlphaProof Code produced machine-checked proofs that each buggy version was insecure. The result concerns those two deliberately reintroduced bugs; it does not establish how the system would perform against every possible defect.
What the reported result adds
WIRED reported that the project took four weeks and that the examined current portions yielded no new vulnerabilities; older versions containing known, fixed vulnerabilities failed formal verification. These findings concern the portions and properties examined. They do not change the boundary between the protocol code tested and the Signal mobile app as a whole.