DeepMind Says AI Has Finally Put Signal’s Code to the Test
DeepMind Says AI Has Finally Put Signal’s Code to the Test
Google DeepMind says it has cleared a long-standing hurdle in cybersecurity: using AI-assisted formal verification to test the actual compiled code behind Signal’s encryption, rather than only the protocol’s design.
The work matters because cryptographic systems can be sound on paper while errors creep in as software is built and compiled. DeepMind said its AlphaProof Code framework paired human experts—who defined and reviewed the security specifications—with agents powered by Gemini Carbon, which generated more than six million lines of Lean proof code. The system then flagged roughly 1% of the codebase for additional human scrutiny.
The company says the four-week effort established key properties including protection of messages before and after a device is temporarily compromised. To test whether the framework could distinguish secure software from flawed software, researchers restored two previously fixed bugs in an isolated version of Signal; in both cases, it produced machine-checked proofs that the altered code was insecure.
The result was publicly framed as a first. In a reposted announcement, DeepMind chief executive Demis Hassabis highlighted his team’s claim that it had proven key security properties of the machine code underpinning Signal, calling it the first formal verification of actual production code for a widely deployed protocol.
Signal’s technology chief, Ehren Kret, cast the development less as a replacement for human judgment than as a major extension of what verification can cover. “Using AI agents to translate live code into valid Lean theorems is a major breakthrough,” he said, arguing that the process gives Signal greater confidence in its privacy promises as AI-driven threats grow.
Axios reported that DeepMind research scientist Swarat Chaudhuri described the broader ambition as reaching “this world of unhackable software,” though the current system is still a research project, not a commercial cybersecurity tool. DeepMind plans to open-source its verification artifacts and explore links to tools such as CodeMender and Project Zero—an early step toward bringing mathematical assurance into everyday software development.
Write a comment