Google DeepMind Formally Verifies Signal Encryption Using Gemini Carbon and AI Agents
DeepMind researchers used AlphaProof Code and Gemini Carbon to formally verify Signal's core encryption protocol, completing a multi-year mathematical proof process in just one month.
Google DeepMind researchers have successfully used AI to formally verify portions of Signal's encryption code, providing mathematical proof that key components of the secure messaging application perform accurately without errors. This achievement marks the first time the implementation of a widely deployed cryptographic protocol has been formally verified by any method, whether manual or automated. Signal's cryptographic architecture is foundational to modern digital communication, as its protocol also powers end-to-end encryption in platforms like WhatsApp and Google Messages.
How DeepMind verified Signal's code
Formal verification relies on automated theorem proving to mathematically demonstrate that software code strictly adheres to its specifications. Historically, creating these formal proofs required human cryptography experts to manually write verification logic over several years, presenting a massive time barrier. To streamline the process, DeepMind researchers built upon prior verification efforts by the Signal Foundation and created an automated workflow powered by AlphaProof Code.
The researchers split the effort between human engineers and autonomous agents. Humans collaborated with AI agents to define the necessary proof targets. Multiple agents running DeepMind's latest Gemini Carbon model then executed the proofs and audited the generated output. The system automatically flagged complex areas requiring human inspection, which reduced the manual workload to just 1% of the codebase. Using this setup, DeepMind completed in a single month what would traditionally take years of manual work.
To confirm that AlphaProof Code could reliably spot defects, the team evaluated the system against two previously resolved Signal bugs. The AI system successfully generated valid proofs while correctly identifying the flaws in the unpatched versions. Although DeepMind did not verify Signal's complete codebase, proving the core protocol gives researchers higher confidence in the application's underlying privacy guarantees.
What it means for developers
For software engineers and security researchers, formal verification has long been viewed as an ideal but resource-prohibitive standard. DeepMind's work demonstrates that AI agents can turn automated theorem proving into a practical tool for real-world software, reducing verification timelines from years to weeks.
This breakthrough has broad implications for software architecture beyond messaging. DeepMind noted that verifying Signal's core cryptography establishes a framework to secure other high-stakes infrastructure, including medical device firmware and power grid control software. Rather than turning the technology into a proprietary consumer product, DeepMind is open-sourcing its code to assist the broader engineering community with verification projects.
As AI-driven testing and verification become standard practices, engineering teams need access to diverse AI models to build and benchmark their tooling. Developers can try top AI models cheaply through one API at https://apixoai.online to integrate models like Gemini Carbon or Claude into their development workflows.
A dual-edged security landscape
The research underscores how AI is acting as both a primary vulnerability vector and a defense mechanism in modern cybersecurity. The year 2026 has been characterized by an increase in AI-related security incidents, many driven by autonomous agents. At the same time, leading AI laboratories are building targeted defensive capabilities, such as Anthropic's Claude Mythos and DeepMind's proof automation tools.
By demonstrating that AI agents can mathematically verify complex protocols, the project provides a scalable path forward for software auditing. As formal verification tools become open and accessible, automated mathematical proofing may soon become a standard requirement for critical codebases across the industry.
Source: Google’s new AI tool just proved Signal encryption really works – here’s how — ZDNet AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
