AI Turns Signal’s Security Proof Into a Four-Week Sprint

Google DeepMind has put a powerful new idea through a real-world test: using agentic AI to prove that critical encryption code works as intended. In just four weeks, researchers formally verified key cryptographic guarantees in the Signal protocol and generated more than 3 million lines of mathematical proofs.

The analysis did not find new vulnerabilities in the parts of the protocol it examined, but its larger message reaches far beyond Signal. If AI can help prove that security protections hold, developers may begin shifting from hunting for bugs to proving that certain weaknesses do not exist.

Signal’s Encryption Gets an Independent Mathematical Check

Google DeepMind used its new AI framework, AlphaProof Code, to mathematically verify a number of Signal’s key cryptographic guarantees. The work serves as independent validation for the cryptography protecting massive quantities of communication across platforms like the Signal app, Google Messages, and WhatsApp.

Google DeepMind said its AI agents, powered by the Gemini 4 Carbon model, verified that the code underpinning Signal’s encryption works as intended. The researchers conducted the effort as a proof-of-concept project, completing it in four weeks—a task Swarat Chaudhuri said would likely have taken humans working alone years.

That speed matters because formal verification has long stood as a kind of holy grail in software security. The process uses mathematics to check whether software functions according to its intended design, including whether it avoids critical vulnerabilities.

Before the AI era, people had to describe the software properties they wanted to prove, translate those descriptions into formulas, and then develop the proof. Agentic AI can cut down the time needed for that work, opening the door to analyses that once demanded far more effort.

“We are very excited by this result, but there is a long distance to go before we are able to scale this to all the software of the world, and we want to do that,” Chaudhuri said.

Old Bugs Put the AI to the Test

The DeepMind analysis did not formally verify the entire Signal mobile app, and the researchers describe the result as one milestone in a much larger process. Still, they tested AlphaProof Code against known weaknesses instead of checking only clean, current software.

Researchers reintroduced two previously fixed bugs into Signal’s code. In both cases, the AI-powered framework generated mathematical proofs showing that the altered software was insecure. When researchers used AlphaProof Code on older Signal versions containing known and long-fixed vulnerabilities, it identified those flaws and showed that the protocol could not be formally verified.

That test gives the system a sharper role than a tool that simply confirms an expected result. It showed that the framework could detect when a security claim failed, as well as produce proofs for the properties that held in the examined portions of the protocol.

Ehren Kret, Signal’s chief technology officer, said, “While prior formal verification focused on smaller portions of the code, Google DeepMind’s approach enables significant expansion.” Kret also said, “As AI-driven security threats grow, rigorously verifying the Signal protocol implementation gives us even greater confidence in our code’s privacy guarantees.”

From Research Project to Software Development

The work remains a research project, not a commercially available cybersecurity product. DeepMind conducted the Signal analysis free of charge, Google declined to share the cost, and Google declined to comment on plans for a commercial formal verification product or service.

The researchers are releasing the specification and proof they developed for the Signal protocol analysis to support other formal verification work. That release could help researchers and developers build on the process instead of starting with an empty page.

DeepMind’s team also points to a human-AI collaboration process as a crucial part of its progress. Agents can contribute and self-correct at almost every phase, while people still provide input and review at key points.

“We don’t want this sort of vibe world where AI is generating the goals and AI is generating the proofs and basically taking tests that are self-written,” Chaudhuri said. “There’s a middle ground.”

Pushmeet Kohli said, “If you can now verify the correctness of these implementations, then you can rest easy.” He added, “You can imagine banks, imagine all those other companies who rely on assurances and who worry about bugs who could now say, ‘I can prove these properties.’”

The potential applications extend beyond messaging encryption. Chaudhuri said the same technology could help developers prove that software is protected against common weaknesses, including memory safety bugs that hackers routinely exploit. DeepMind is also exploring ways to integrate the verification technology into AI coding agents and software development workflows, allowing developers to verify security protections as they build and update software.

The team is exploring how to combine the research with Google’s existing vulnerability-hunting and patching tools, including Project Zero and CodeMender. For now, the Signal proof-of-concept marks an important step toward software that can demonstrate its security properties with mathematics rather than assurances alone.

“You can finally get to this world of unhackable software,” Chaudhuri said. The path to that world remains long, but four weeks and more than 3 million lines of proofs show how quickly AI may move formal verification forward.

Stay connected via Google News

Follow us for the latest AI News