Google DeepMind Uses AI Agents to Formally Verify Signal's Post-Quantum Cryptography
Google DeepMind researchers have employed an AI-driven framework to formally verify the security of Signal's post-quantum cryptographic handshake and ratcheting mechanisms.
In a significant advancement for AI-assisted cybersecurity, Google DeepMind researchers have successfully used their AlphaProof Code framework to formally verify critical components of Signal's secure messaging protocol. This marks a milestone in applying artificial intelligence to rigorously prove the correctness of widely-used cryptographic implementations, specifically targeting the post-quantum PQXDH handshake and the Triple Ratchet algorithm.
The AlphaProof Code system operates as a human-guided, multi-agent AI framework. It begins with human input to define a specification of the security goals, then leverages AI agents to generate millions of lines of formal proof code. This code is then checked against the initial specification by a formal verification kernel, ensuring that the implementation adheres to the intended security properties. In this instance, the AI agents generated approximately 6.5 million lines of Lean code to validate Signal's libsignal implementation, confirming its forward secrecy, authentication, and post-compromise security.
This effort is part of a broader initiative, including a challenge issued by the Beneficial AI Foundation, to formally verify the entire Signal protocol and application using formal methods. While the verification is theoretical and assumes perfect cryptographic primitives rather than guaranteeing computational secrecy of data in transit, it provides a strong assertion of the protocol's design correctness at the implementation level. The researchers emphasize that this is a step towards scaling formal verification, a process traditionally requiring immense human effort.
Swarat Chaudhuri, a research scientist at Google DeepMind, highlighted the shift in focus within formal methods research. "In formal methods, we always had this belief that verification is the hard bit," Chaudhuri stated. "Now we are able to do that. So now the question is what are we verifying? … That process is fundamentally human-in-the-loop, but we devised a process of trying to scale it up to the extent possible with AI."
Traditionally, formal verification projects have been resource-intensive, with efforts like verifying the seL4 microkernel taking 20 person-years. AlphaProof Code aims to drastically reduce this burden. Although the Signal verification took a team of five people four weeks and significant computational resources, it represents a substantial acceleration compared to manual methods. The researchers believe this approach can be applied to a much wider range of software.
Looking ahead, Chaudhuri indicated that the next frontier lies in AI's role in the specification process itself, not just proof generation. The team plans to apply similar technologies, like the related AlphaProof Nexus project focused on mathematics, to assist in the software design process more broadly. "We really believe that the next frontier is not going to be proof but specification in this kind of research," he added.
The implications of this work extend beyond Signal. As AI capabilities in offensive security grow, the need for robust, formally verified defensive technologies becomes paramount. By demonstrating the scalability of AI-assisted formal verification, Google DeepMind is paving the way for more secure software development across the industry, aiming to guard against the increasing sophistication of cyber threats.