DeepMind details AlphaProof Code: agentic framework that formally verified Signal's protocol
sundarpichai · x · 2026-11-11
Google DeepMind published the technical details of AlphaProof Code, a multi-LLM agentic framework that split formal verification like a research team.
- Three-stage flow: human experts + AI translate Signal's security goals into formal specs (each human-reviewed); Signal's Rust code is compiled into an executable Lean processor model; a Gemini Carbon-powered prover team generates proofs with a gatekeeper filtering rejects
- First formal verification of a widely deployed cryptographic protocol implementation by any method: 6M+ lines of Lean proofs, 1% human review
- Results to be open sourced; goal is shifting the advantage to defenders
More from Safety
- Ought sponsors $500k–$5M Epistemics Fund for community notes and provenance tools — NathanpmYoung · 2026-11-11
- Are theorem prover bugs a security risk? Debate ignites as LLMs write proofs — OwariDa · 2026-11-11
- Railway's 'novel-length' postmortem of its Oct 21 security attack earns rare praise — kingslyj · 2026-11-11
- Databricks engineer: other AI inference providers' security breaches vindicate our paranoia — Yuchenj_UW · 2026-11-11
- "Regulation is a wet pool noodle against a nuke": dev argues LLMs will be countered soon — fkasummer · 2026-11-11
- AI safety researcher: make systems provably safe before release, no slowdown needed — ghadfield · 2026-11-11