2QR / JOURNAL✳

The development

On November 10, Google DeepMind introduced AlphaProof Code, a framework that combines human guidance with AI agents to verify security properties of software. The team applied it to the code underlying the Signal protocol, which protects private communications in several messaging products.

The researchers say their system generated more than six million lines of proof code in Lean, a tool that checks mathematical proofs. Humans reviewed the formal specifications that set out what the software was supposed to do; automated agents then worked on proving that the implementation satisfied those specifications.

Why this matters for software security

Most software testing checks selected examples. Formal verification takes a different approach: it tries to prove a precisely defined property across the cases covered by a mathematical model. If the specification is incomplete or an important component lies outside the proof, the result does not automatically cover that gap.

DeepMind also reports that the system identified insecurity when two previously fixed bugs were reintroduced into an isolated version of the code. That is encouraging evidence about this experiment, not a promise that any messaging app or computer is impossible to compromise.

What comes next

The practical challenge is making rigorous proofs cheaper to produce, maintain, and review as software changes. If that becomes easier, teams responsible for sensitive systems could use AI to assist security specialists rather than relying on tests alone. DeepMind says it is opening its verification artifacts for further use and scrutiny.

Source

Google DeepMind’s research announcement

Cover image: AI-generated conceptual illustration.

✳
← Back to all stories