- Google DeepMind has formally verified the security of the Signal protocol, which underpins encryption in Signal, WhatsApp, and Google Messages.
- The verification was achieved using AlphaProof Code, a multi-agent framework combining AI-driven analysis with human oversight.
- This landmark development validates Signal’s robust cryptographic guarantees and showcases AI’s potential to streamline formal verification processes.
Google DeepMind Verifies Signal Protocol Security with AI
Researchers at Google DeepMind have successfully completed a formal verification of Signal’s messaging security, marking a significant milestone as the first instance where AI has assisted in devising a proof linking protocol-level security guarantees to a widely used cryptographic implementation. The Signal protocol, long regarded as the gold standard for end-to-end encrypted messaging, is not only utilized in the Signal app but also by WhatsApp and Google Messages.
AlphaProof Code: A Hybrid Approach to Verification
The achievement was made possible through AlphaProof Code, a multi-agent framework designed for formal software specification and verification. AlphaProof Code employs a human-guided agent that writes a Lean specification, assesses machine code semantics, and crafts an AI-authored proof that the Lean kernel checks against the specification. In this project, AlphaProof Code verified Signal’s post-quantum PQXDH handshake and Triple Ratchet, thereby validating the design of libsignal’s forward secrecy, authentication, post-compromise security, and handshake secrecy.
The AlphaProof Code process involves converting Signal’s security objectives into a human-reviewed mathematical specification. Based on this specification, Gemini Carbon agents were assigned to generate Lean code proving the correctness of the libsignal implementation. After generating 6.5 million lines of Lean code, the agents performed an auditing pass, flagging sections that required human review. While the result is theoretical, assuming perfect cryptography and not guaranteeing the computational secrecy of bytes, it represents a meaningful assertion of correctness.
Reducing Verification Costs and Future Implications
Swarat Chaudhuri, a research scientist at Google DeepMind and a professor at the University of Texas at Austin, noted that AlphaProof Code is fundamentally built for human-AI collaboration. Traditionally, proving theorems involves mathematicians formulating well-defined problems for AI agents to solve. In this case, the researchers had to determine what needed to be proven. “In formal methods, we always had this belief that verification is the hard bit,” Chaudhuri stated. “There’s thirty or forty years of work on how to scale up verification. 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.”
As Chaudhuri and his colleagues highlight in their research paper, “AlphaProof Code: Verifying Signal’s Security at the Machine-Code Level,” formal verification historically demanded immense effort. For instance, formally verifying the seL4 microkernel required 20 person-years, while developing the verified CompCert C compiler took six years.
The Google/DeepMind team-comprising Swarat Chaudhuri, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger, Karn Seth, Jonathan Katz, Charles Sutton, and Pushmeet Kohli-believes AlphaProof Code can significantly reduce the cost of formal verification. Although their work on Signal still took four weeks for a team of five people and “quite a few tokens,” it represents a substantial leap forward. “The dream is definitely to scale this up,” Chaudhuri commented, referencing the rise in offensive AI capabilities and the prevalence of insecure systems globally. The next frontier, he suggests, is to apply this technology during the software design process, not just to cryptographic protocols but to code in general. This aligns with broader industry trends, such as the emergence of AI-focused operating systems like Kora, indicating a growing emphasis on integrating AI into fundamental software development.
| Parameter | Traditional Verification | AlphaProof Code (AI-aided) |
|---|---|---|
| Time (seL4 example) | 20 person-years | Significantly reduced |
| Time (CompCert C example) | 6 years | Significantly reduced |
| Time (Signal verification) | Not specified | 4 weeks (5-person team) |
| Human involvement | High, problem formulation | High, but AI-scaled assistance |
| Scalability | Limited | High, aims for general code verification |
For users, this formal verification provides enhanced assurance regarding the robustness of end-to-end encryption in widely used applications like Signal, WhatsApp, and Google Messages, reinforcing trust in their communication privacy. For developers and the software industry, it signifies a breakthrough in AI-aided formal verification, offering a pathway to substantially reduce the cost and time required to secure critical systems. This advancement could lead to the development of much more secure software across all layers, from cryptographic protocols to general codebases, improving overall digital security.
I’ve been using Signal for years, precisely because of its reputation for robust encryption. This DeepMind verification, especially with AI assistance, really solidifies my trust. I always tell friends to switch, and the end-to-end encryption has always worked flawlessly in my experience. The only minor frustration is getting everyone on board; it’s a hurdle. For new users, my tip is to set up disappearing messages for an extra layer of privacy, it’s a game-changer once you get used to it.