DeepMind uses AlphaProof Code to formally verify Signal's protocol
Original title: Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
Gemini Carbon agents produced over six million lines of Lean in four weeks and flagged two reintroduced bugs.
Google DeepMind says researchers used AlphaProof Code, a multi-LLM agentic framework, to formally verify security properties of the Signal protocol implementation. Human experts set specifications, Gemini Carbon agents generated Lean proofs, and people reviewed about one percent of the codebase. The lab says the four-week effort is the first verification of this widely deployed implementation down to compiled code, and it is open-sourcing the artifacts. Reintroduced copies of two previously fixed bugs were proved insecure.
Unconfirmed: The first-of-its-kind verification claim comes from Google DeepMind and is not independently confirmed in the post.
- Proofs cover message protection before and after temporary device compromise.
- DeepMind is open-sourcing the formal verification artifacts and a technical paper.
- Signal researchers gave feedback on the methodology.
- CodeMender is cited only as a related automated defense agent.
- Products
- AlphaProof CodeSignal
- Organizations
- Google DeepMind
- AI models
- Gemini Carbon
From the original article
4 min read at deepmind.google
AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal’s core encryption.
By Pushmeet Kohli and Swarat Chaudhuri · deepmind.google
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.