AI cuts encryption check from years to weeks—DeepMind verifies Signal

DeepMind checks Signal’s encryption with AI

Researchers at Google DeepMind used an AI tool called AlphaProof Code to check Signal’s encryption protocol. In four weeks, they mathematically confirmed that core parts of the code work exactly as designed—no errors, no backdoors. That same protocol secures messages on WhatsApp and Google Messages.

How AI and engineers worked together

AlphaProof Code splits the task into smaller proofs. Human engineers set the goals and step in when the AI hits dead ends. Swarat Chaudhuri, a DeepMind research scientist, told WIRED that this back-and-forth cut the time from years to weeks.

Swarat Chaudhuri, another DeepMind research scientist on the project, describes an "iterative interaction" where AI proposes steps and engineers adjust the direction when needed. No single model handles everything; while AlphaProof Code handles mathematical proofs, a separate large language model called Gemini Carbon translates plain-text descriptions of Signal’s protocol into formal logic.

What the team checked—and what it missed

The team focused on the core Signal protocol—the part that handles key exchanges and message encryption. That protocol is identical across Signal, WhatsApp, and Google Messages. While the verification didn’t cover the entire Signal mobile app, it did confirm the absence of vulnerabilities in the sections it reviewed.

To test the system, DeepMind fed AlphaProof Code two old Signal bugs that had already been patched. The AI correctly flagged the flawed versions and accepted the corrected ones.

No product yet—just open-source tools and doubts

DeepMind ran the project as a free research proof-of-concept; Google won’t say how much it cost or whether a commercial product is coming. Instead, the team is open-sourcing the specification and proofs they built for Signal. That move could help other researchers apply formal verification to new projects.

So far, the industry remains skeptical. Ehren Kret, Signal’s chief technology officer, told WIRED that AI-driven verification is "magical when it works, but breaking it down into a scalable, cost-effective process is still an unsolved problem." Pushmeet Kohli, DeepMind’s chief scientist, acknowledges the uncertainty but calls the work "a step toward improving our security infrastructure."

Why this matters: speeding up software guarantees

Formal verification has sat on a pedestal for decades: a mathematical guarantee that software behaves exactly as promised. The catch? Writing those proofs by hand can take teams of PhDs years. DeepMind’s AI doesn’t eliminate the need for human oversight, but it does automate the most tedious parts—translating code into logic, generating draft proofs, and flagging contradictions.

The Signal project shows what’s possible when AI tackles a focused, high-value target. Scale it up, and the same approach could help banks, hospitals, or power grids prove their systems are truly secure.

AI’s growing role in security and verification

Last year Red Hat OpenShift 5 added AI-powered troubleshooting for Kubernetes clusters. That update and DeepMind’s work are part of the same larger shift: AI handling the repetitive, detail-heavy work that slows down software verification and security audits.

For encryption protocols like Signal’s, that shift could mean faster responses to new threats and quicker rollouts of fixes across messaging platforms.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

AI-powered verification could cut the time and cost of securing critical software—if the tech industry figures out how to scale it.

Key takeaways

• DeepMind’s AI slashed verification time from years to weeks, but didn’t check the entire Signal app.
• The same math that secures Signal also protects WhatsApp and Google Messages.
• AlphaProof Code caught known flaws in old Signal bugs, proving its accuracy.
• DeepMind shared its tools for free, yet cost and scalability doubts persist.

Related on Lazyfounder

• Google DeepMind
• Google
• Red Hat OpenShift 5 brings AI-driven Kubernetes and quantum-ready security