“DeepMind's AlphaProof Code agents generated three million lines of Lean proofs to formally verify the core encryption underpinning Signal's messaging app. This marks a significant leap in AI-assisted formal verification, demonstrating that AI can operate at the scale and depth required for real-world, security-critical software. The achievement suggests that provably correct software for high-stakes systems may be within practical reach.”
Key Takeaways
- AlphaProof Code agents produced three million lines of Lean proofs to verify Signal's core encryption library.
- This is DeepMind's first application of AlphaProof-style agents to real-world production software rather than mathematical benchmarks.
- Formal verification at this scale was previously impractical without AI assistance, representing a step-change in software assurance capability.
DeepMind's AlphaProof Code wrote three million lines of formal proofs to verify Signal's encryption.
Part of the 10 November 2026 AI news digest
trending_upWhy It Matters
Formal verification has long been considered too expensive and labour-intensive for widespread adoption in production software, limiting its use to aerospace and hardware design. If AI agents can automate proof-writing at million-line scale, security-critical software — from messaging apps to financial infrastructure — could achieve mathematically guaranteed correctness. This could fundamentally shift how the industry approaches software auditing, reducing reliance on penetration testing and code review alone. Regulators pushing for higher assurance in AI systems and critical infrastructure should take note, as this tooling could become a compliance mechanism. The next question is whether AlphaProof Code can generalise beyond encryption to other complex software domains.
FAQ
What is Lean, and why does it matter for this breakthrough?
Lean is a formal proof assistant and programming language used to write mathematically verifiable proofs about software behaviour. Using Lean means the correctness of Signal's encryption isn't just tested but mathematically proven, eliminating entire classes of bugs and vulnerabilities.
Does this mean Signal's encryption is now completely bug-free?
Formal verification proves that the code meets its specified mathematical properties, which is a much stronger guarantee than traditional testing. However, it only covers what the specification defines — flaws in the specification itself or in underlying hardware remain outside the scope of this verification.
How does AlphaProof Code differ from DeepMind's original AlphaProof?
The original AlphaProof was designed to solve olympiad-level mathematics problems using formal reasoning in Lean. AlphaProof Code extends this capability to software verification, applying the same proof-generation approach to real production codebases rather than abstract mathematical challenges.



