Google DeepMind uses AI to mathematically verify Signal encryption

Automated formal verification

Google DeepMind recently utilized a new agentic AI framework called AlphaProof Code to examine the security of Signal. This project focused on the messaging app's encryption protocol, which is designed to keep private conversations secure from outside observers. The goal was to ensure the software actually provides the privacy protections it claims to offer.

The team employed a method known as formal verification. This process involves using mathematical proofs to demonstrate that a piece of software works exactly as intended. Because it relies on rigorous math rather than just testing for bugs, it is considered one of the highest standards for software security available today.

Historically, formal verification has been a very difficult and slow process. It usually requires large teams of human experts to spend a significant amount of time writing out every single proof. By using AI, Google DeepMind aims to make this high-level security check more accessible for complex systems.

How the AI performed the check

To conduct the verification, the researchers used AI agents powered by Gemini. These agents were tasked with generating millions of lines of mathematical proofs to scrutinize the Signal Protocol. This allowed the system to systematically check the code for any logical errors or security gaps that might exist in the underlying math.

As reported by The SOFX Report, the AI was able to complete this massive undertaking in just four weeks. Swarat Chaudhuri, a research scientist at Google DeepMind, noted that the same work would likely have taken humans several years without the help of AI. This speed represents a significant leap in how quickly developers can audit complex encryption.

The AI did not just look for new problems; it also acted as a diagnostic tool for past issues. During the process, the framework correctly identified specific security problems that experts had previously found and patched in older versions of the software. This confirms the AI's ability to recognize known patterns of vulnerability within a codebase.

Impact on Signal and privacy

The results of this study provide a new layer of confidence for users of the Signal app. Because the encryption was mathematically verified, users can have greater certainty that their data remains private. Signal's chief technology officer, Ehren Kret, said the work provides greater confidence that the app’s encryption delivers the privacy protections it promises.

While the AI did not uncover any new vulnerabilities in the specific parts of the protocol it examined, the process itself is a milestone. It marks the first time that the actual code of a widely used encryption protocol has been formally verified using AI or any other method. This sets a new precedent for how public-facing security tools are audited.

The success of this project suggests that the Signal Protocol is robust. However, the project also highlights the importance of verifying the specific parts of the code that are currently in use. By proving the math behind the code, the researchers have established a baseline of trust for the app's core security features.

Broader applications for AI security

The implications of this project extend far beyond a single messaging app. Google DeepMind suggested that this technology could make advanced software security checks faster and less expensive for other industries. This could eventually allow for the verification of critical systems that require extreme reliability, such as power grids or medical devices.

The ability to quickly verify that software behaves correctly is vital for infrastructure that cannot afford to fail. If AI can handle the heavy lifting of mathematical proofs, it could lead to a new standard for how we deploy high-stakes technology. This would move the needle from reactive bug-hunting to proactive mathematical certainty.

Other organizations are already showing interest in this specific niche of AI development. Companies like Anthropic and startups such as Theorem, &Truth and Harmonic are exploring AI-powered mathematical verification. This indicates a growing trend toward using AI to identify and prevent software vulnerabilities before they can be exploited by bad actors.

Challenges and future outlook

Despite the success of the AlphaProof Code framework, there are still practical hurdles to overcome. Researchers have noted that questions remain regarding the cost of running these large-scale AI models. There are also concerns about the scalability of the process when applied to even larger, more complex software ecosystems than the Signal Protocol.

Furthermore, the project underscores the continued need for human oversight in the security field. While the AI can generate millions of lines of proofs, humans are still required to interpret the results and ensure the framework is being applied correctly. The goal is to augment human expertise rather than replace it entirely in high-stakes security environments.

Moving forward, the industry will likely watch how these tools evolve to handle diverse types of software. As more companies begin to experiment with AI-powered verification, the focus will shift to how these tools can be standardized. The ultimate goal is to create a world where software security is not just a best effort, but a mathematically proven guarantee.