SIAgentNews.com
← All news

Google DeepMind uses AI to mathematically prove Signal's security

Researchers used a new agentic framework to formally verify the actual code of the Signal protocol, a task that usually takes human experts years.

· First reported by Google DeepMind

four weekstime taken to verify the protocol

The short version

  • The project used AlphaProof Code, a framework combining human expertise with AI agents to perform formal verification.
  • The system generated over six million lines of Lean, a language used to mathematically prove code accuracy.
  • The verification process took just four weeks, whereas similar tasks historically take expert teams years.
  • The team proved that the compiled code protects messages even if a device is temporarily compromised.
  • Researchers confirmed the system's accuracy by successfully identifying two real, previously fixed bugs in the code.

Why it matters

This proves that AI can handle high-level security tasks that were previously too complex for automation. It creates a blueprint for verifying other critical infrastructure, like power grids and medical devices, with mathematical certainty.

A new standard for code safety

Google DeepMind has reached a milestone in software security by using AI to mathematically prove the security properties of the Signal protocol. This protocol is the underlying technology that protects private conversations for billions of people using apps like Signal and Google Messages. While many systems claim to be secure, this project provides a formal mathematical proof that the code actually works as intended.

Formal verification is considered the highest standard of security because it proves that code runs correctly without any flaws. Historically, this process has been extremely labor-intensive and difficult to achieve for complex systems. This work represents a first for the industry, as it is the initial instance of AI being utilized to provide a mathematical proof for a cryptographic protocol used at such a large scale.

The team specifically focused on the implementation level of the protocol. Most previous verification efforts stopped at the design level, which only confirms the theory behind the security. By verifying the actual compiled code, the researchers bridged the gap where real-world vulnerabilities often hide between a design and its execution.

How AlphaProof Code works

The researchers developed a new agentic framework called AlphaProof Code to handle this massive task. This framework splits the work similarly to how a human research team would operate. Human experts collaborated with a mix of models to translate Signal's security goals into formal mathematical specifications, which the humans then reviewed.

Once the goals were set, agents powered by Gemini Carbon performed the heavy lifting of the proofs. These agents generated over six million lines of Lean, which is a computer language designed to mathematically prove the accuracy of code. Because Lean automatically checks the rest of the work, the human experts only had to review one percent of the codebase.

This hybrid approach allowed the team to focus human judgment where it mattered most. The AI agents handled the vast majority of the repetitive and complex theorem proving. This combination of human oversight and AI scale allowed the team to complete the verification in just four weeks.

Proving protection against threats

The verification confirmed that the Signal protocol fulfills key security properties even under specific conditions. For example, the proof showed that messages remain protected before and after a device is temporarily compromised. This level of assurance is critical because cryptographic protocols underpin almost all digital communication, including online banking and web browsing.

To ensure the system could actually detect errors, the researchers reintroduced two real, previously fixed bugs into an isolated version of the Signal code. In both instances, the AlphaProof Code system produced a machine-checked proof showing that the buggy version was insecure. This test confirmed that the framework can distinguish between correct code and flawed code.

The project was a collaboration involving researchers from Google DeepMind and feedback from the Signal team. The researchers are now open-sourcing their formal verification artifacts for the broader community to use. This move aims to make these high-level security tools more accessible to other developers and organizations.

Scaling security for the future

This achievement shows how artificial intelligence can provide high-level security guarantees to various sectors, including governments and critical infrastructure. As AI-driven cyber threats grow, having the core Signal protocol rigorously verified gives more confidence in privacy guarantees. The success of this project lays the groundwork for securing other vital systems that require high reliability.

The researchers suggest that these techniques could eventually be used to verify medical device firmware or power grid software. These systems are often too complex for manual verification but are too important to leave to chance. By using automated theorem proving, developers can create software with a mathematical proof that its protections actually work.

The future of this technology involves pairing formal verification with automated defense agents like CodeMender. This combination could help counter the speed and scale of modern cyber threats. While scaling these techniques requires continued research, the project proves that AI can handle the most demanding security challenges currently facing the industry.

What to watch

  • How other major communication protocols like WhatsApp or iMessage will adopt formal verification.
  • The expansion of AlphaProof Code to verify hardware-level firmware for medical and industrial devices.

Try it

You can look for the technical paper on Google DeepMind's website to see the specific mathematical specifications they used.

Sources and links

First reported by Google DeepMind. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news