Google DeepMind formally verifies Signal protocol using AI-powered AlphaProof Code
Google DeepMind used agentic AI to produce a machine-checked mathematical proof of key Signal protocol cryptographic guarantees, generating over six million lines of Lean proof code in four weeks.
Summary
Google DeepMind researchers have completed a formal verification of core components of the Signal protocol using a new AI framework called AlphaProof Code, producing what they describe as a landmark result for both cryptography assurance and automated reasoning. The four-week proof-of-concept project translated Signal's security goals into formal mathematical specifications with human expert review, then used Gemini Carbon-powered agents to generate and audit over six million lines of Lean proof code. The analysis did not uncover new vulnerabilities in current Signal implementations, but successfully identified known historical flaws when applied to older protocol versions, demonstrating the technique's validity.
The work provides independent validation for the cryptography protecting communications across Signal, Google Messages, and WhatsApp. DeepMind is releasing the specification and proof artifacts to support further formal verification efforts in the cryptographic community.
How AlphaProof Code verified Signal
- Formal specification pipeline: Researchers collaborated with multiple AI models to translate Signal's security goals into formal mathematical specifications, with human experts reviewing each translation step.
- Lean proof generation: Gemini Carbon-powered agents generated over six million lines of Lean proof code, the dependently-typed language used for machine-checked mathematical verification of software correctness.
- Automated proof auditing: The AI agents audited their own proofs and flagged approximately one percent of the codebase for mandatory human review, with Lean's automated checking handling the remaining ninety-nine percent.
- Regression validation: When applied to older Signal versions containing known and previously fixed vulnerabilities, the system correctly identified these flaws and demonstrated that those protocol versions could not pass formal verification.
- Open artifacts: DeepMind is releasing the formal specification and complete proof developed during the analysis to support broader adoption of AI-assisted formal verification in cryptography.
Why it matters
This demonstration matters for the open-source security community because it shows a viable path to reducing the cost and time of formal verification, which has historically been prohibitively expensive for most projects. If AI-assisted proof generation becomes accessible beyond well-resourced research labs, critical open-source infrastructure like TLS libraries, VPN protocols, and messaging implementations could receive stronger mathematical assurance. The release of proof artifacts also establishes a reproducible baseline that other researchers can extend or challenge, advancing the state of publicly auditable cryptography.
Original reference: www.schneier.com →