Google DeepMind says it has used artificial intelligence agents to formally verify security properties in the compiled code behind Signal’s messaging protocol. The development could make rigorous software checks less expensive, although it does not mean the messaging application is completely secure.
Published on November 10, 2026, the research uses AlphaProof Code. AI agents helped produce mathematical proofs, which the Lean proof system checked. Google DeepMind+1
What DeepMind Actually Verified
The research examined libsignal, the Signal Protocol implementation, focusing on its PQXDH post-quantum handshake and Triple Ratchet. The paper describes these as the first proofs linking protocol-level security guarantees to an unmodified, widely deployed cryptographic implementation. Its Signal analysis is limited to one-to-one sessions; group messaging remains outside the study.
DeepMind says it proved specified properties including authentication, forward secrecy, post-compromise security and handshake secrecy. These protections help authenticate contacts and limit damage if a device or key is compromised.
Researchers also reintroduced two previously fixed bugs into an isolated version of the code. The system produced machine-checked proofs showing that the altered versions failed the required security claims. Google DeepMind+1
Also Read: Google DeepMind Introduces Gemini 4 Carbon AI Model With Advanced Cybersecurity Capabilities
How AI and Human Experts Worked Together
AlphaProof Code combines human judgement with AI agents. Researchers used AI models to translate Signal’s security goals into formal specifications written in Lean. Human experts reviewed the specifications before agents powered by Gemini Carbon generated proofs.
Lean checked the proofs independently. Unlike a persuasive AI explanation, a machine-checked proof must follow from its formal assumptions.
DeepMind reports that the project generated approximately 6.5 million lines of Lean code and took four weeks. The agents flagged about one per cent of the codebase for additional human review, according to the company. They do not mean every part of Signal was examined. Google DeepMind+1
What the Findings Mean, and What They Do Not
The announcement is not evidence that Signal is unhackable. It covers defined properties of particular cryptographic components and their compiled implementation. It does not establish that the entire application, phones, operating systems, user accounts or surrounding infrastructure are free from vulnerabilities.
The research also relies on trusted specifications, assumptions about cryptographic primitives and a formal model of the processor. A proof cannot guarantee protection against threats excluded from its specifications or assumptions. Google DeepMind+1
For Nigerian banks, telecommunications companies and public digital-service providers, the research highlights a possible direction for strengthening software assurance. However, AlphaProof Code is not an off-the-shelf security product. Applying the approach elsewhere requires specialist expertise, carefully defined requirements and further work to scale it.
Back Story: Why Proving Software Security Is Difficult
Software security has long relied on code reviews, testing and vulnerability assessments. However, tests examine selected conditions, not every possible execution.
Formal verification expresses required software behaviour as mathematics and checks whether an implementation satisfies those requirements. Applying it to large systems has required substantial specialist effort.
DeepMind’s contribution is to use AI agents to help with that work while retaining human review and an independent proof checker. The company has released verification materials for researchers to examine. The next challenge is determining whether the approach can be independently scrutinised and adapted to other systems without weakening the guarantees. Google DeepMind+1
