SIAgentNews.com
← All news

Google DeepMind uses AI agents to verify Signal's messaging security

Researchers used a multi-agent framework to create the first AI-assisted proof linking protocol security to a major cryptographic implementation.

· First reported by The Register

6.5 millionlines of Lean code generated by agents

The short version

  • The team verified Signal's post-quantum PQXDH handshake and Triple Ratchet protocols.
  • AlphaProof Code generated 6.5 million lines of Lean code to prove the implementation's correctness.
  • The process combined human-guided specifications with AI-authored proofs and auditing.
  • The project took four weeks for a team of five people to complete.
  • The verification confirms the design of forward secrecy, authentication, and handshake secrecy.

Why it matters

This demonstrates that AI can significantly reduce the time and cost required to prove that complex software is secure. It paves the way for verifying large-scale systems that were previously too complex for human mathematicians to audit alone.

Verifying Signal's Security

Google DeepMind researchers have completed a formal verification of the security protocols used by the Signal messaging app. This work represents a significant milestone because it is the first time AI has helped create a proof linking protocol-level security guarantees to a widely used cryptographic implementation.

The team specifically verified Signal’s post-quantum PQXDH handshake and the Triple Ratchet. These components are essential for maintaining the app's security standards, including forward secrecy, authentication, and post-compromise security. By completing this task, the researchers validated the underlying design of the libsignal library.

The project was prompted by a challenge from the Beneficial AI Foundation, which called for the formal verification of the Signal protocol and app using Lean. Lean is a functional programming language and theorem prover often used by mathematicians to write formal proofs of software correctness.

How AlphaProof Code Works

The researchers utilized a multi-agent framework called AlphaProof Code to achieve these results. This system relies on a human-guided agent that writes a Lean specification and assesses the semantics of machine code. It then crafts an agent-authored proof that the Lean kernel checks against that specific specification.

In this specific workflow, the team worked with various models to translate Signal's high-level security goals into a human-reviewed mathematical specification. Once the specification was set, they assigned Gemini Carbon agents to generate the actual Lean code needed to prove the correctness of the libsignal implementation.

The AI agents produced 6.5 million lines of Lean code during the process. After generating this massive amount of code, the agents performed an auditing pass to flag specific sections that required a human to review and finalize.

Scaling Formal Verification

Formal verification is the process of using mathematical proofs to ensure a system behaves exactly as intended. Traditionally, this process has been incredibly labor-intensive and slow. For example, verifying the seL4 microkernel took 20 person-years, while developing the verified CompCert C compiler took six years.

The Google DeepMind team believes AlphaProof Code can drastically reduce the cost and time required for these types of proofs. While their specific work on Signal took four weeks for a team of five people, they view this as a scalable model. They are moving away from a model where humans write every theorem to one where AI handles the heavy lifting of proof generation.

Swarat Chaudhuri, a research scientist at Google DeepMind, noted that the system is fundamentally built for human-AI collaboration. He explained that while verification used to be the hardest part of the process, the goal is now to figure out what specifically needs to be verified at scale.

Theoretical vs. Practical Security

It is important to understand the scope of this specific verification. The result is theoretical in nature because it assumes perfect cryptography. It does not provide a guarantee of the computational secrecy of individual bytes, but it does offer a meaningful assertion of correctness for the protocol design.

The research highlights a shift in how security researchers might approach software in the future. Instead of relying solely on testing, which might miss certain edge cases, formal verification provides a mathematical certainty that the code follows the intended rules. This is particularly useful for high-stakes cryptographic protocols like those used by Signal.

The team is looking toward a future where this technology can be applied to code at large. They want to move beyond just cryptographic protocols to help guard against exploits in insecure systems across the world. This involves figuring out how to write and specify the properties that need to be guarded against in general software.

The Next Frontier for AI

Looking ahead, the researchers want to apply this technology during the initial software design process. They believe the next frontier in this research is not just the proof itself, but the specification of what the software is supposed to do. This means using AI to help define the rules that the software must follow from the very beginning.

The team is drawing inspiration from other projects like AlphaProof Nexus, which focuses on mathematics. By applying similar logic to software, they hope to create a standard way to verify complex systems. They view the Signal project as a way to gain insights into a very hard problem without claiming to have the final word on the technology.

As offensive AI capabilities rise, the need for robust verification becomes more urgent. The researchers aim to eventually apply these methods to code all over the world to identify and mitigate vulnerabilities before they can be exploited by malicious actors.

What to watch

  • How AlphaProof Code will be adapted for non-cryptographic software.
  • Whether more major messaging apps will adopt formal verification for their protocols.
  • The development of automated tools to help humans write formal specifications.

Try it

Check the Signal app's security documentation to see how they describe 'Forward Secrecy' and 'Post-Compromise Security'.

Sources and links

First reported by The Register. 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