Google DeepMind researchers have used agentic artificial intelligence to mathematically prove the security of the Signal protocol. This development marks a significant milestone for both cryptography and AI, demonstrating that automated systems can validate the mathematical foundations of software that protects billions of communications.
The project, which took four weeks to complete, serves as a proof-of-concept for a new AI framework called AlphaProof Code. The work provides independent validation of the cryptographic standards that secure messages on platforms including Signal, Google Messages, and WhatsApp. It also highlights the growing capability of AI to handle complex formal verification tasks, a process that traditionally requires extensive human effort.
Mathematical Proof of Cryptographic Guarantees
Formal verification uses mathematics to check that software functions exactly as intended. Before the advent of advanced AI, this process was incredibly time-consuming. Researchers had to manually describe software characteristics, translate them into formulas, and develop proofs. Agentic AI significantly reduces the time required for these steps, enabling analyses that were previously impractical.
The DeepMind analysis did not discover any new vulnerabilities in the portions of the Signal protocol it examined. However, the system demonstrated its reliability by correctly identifying known vulnerabilities in older versions of the protocol. When researchers applied AlphaProof Code to these legacy versions, the AI successfully showed that the code could not be formally verified, confirming the presence of flaws that had since been fixed.
Researchers are releasing the specification and proofs developed during this analysis to assist other formal verification efforts. This transparency aims to help the broader security community build more robust verification tools.
Human-AI Collaboration in Security
While the speed of the project was notable, the researchers emphasized that they had a head start due to prior formal verification work conducted by the Signal Foundation. The DeepMind approach builds on these earlier efforts, allowing for a significant expansion of what can be verified.
Ehren Kret, the chief technology officer at Signal, noted that while previous verification focused on smaller code segments, the new approach enables broader analysis. He stated that rigorously verifying the Signal protocol implementation provides greater confidence in the platform’s privacy guarantees, especially as AI-driven security threats evolve.
Swarat Chaudhuri, a DeepMind research scientist, described the work as a “human-AI collaboration process.” He explained that the system allows agents to contribute and self-correct at almost every phase while still seeking human input at key junctures. Chaudhuri warned against a scenario where AI generates both the goals and the proofs without oversight. He emphasized that there is a middle ground between full automation and manual verification.
Scaling Verification for Critical Infrastructure
The ability to formally verify code is often viewed as a holy grail in software security. For critical services used in high-stakes environments, the capacity to prove that implementations are correct would provide substantial assurance. Pushmeet Kohli, chief scientist at Google Cloud and vice president of research at DeepMind, suggested that industries such as banking could benefit from such assurances. He noted that companies relying on bug-free software could potentially prove their properties with greater certainty.
However, questions remain about whether AI-automated formal verification can scale to all software. The DeepMind analysis was conducted as a free research project, and Google declined to share the specific costs involved. The company also did not comment on plans to release a commercial formal verification product.
The scope of this analysis was limited. It did not formally verify the entire Signal mobile application. Researchers stress that this finding is just one step in a larger process. Swarat Chaudhuri stated that while the result is exciting, there is a long distance to go before this technology can be scaled to all software in the world.
Source: Wired
