AlphaProof Code: AI agents formally verify Signal's crypto code in four weeks
Google DeepMind's multi-LLM agentic framework for formal verification: human experts decide what to prove, then Gemini Carbon agents write machine-checked Lean proofs down to compiled machine code.
Why it matters
First AI formal verification of a widely deployed cryptographic protocol's actual code: 6.5 million lines of machine-checked Lean proof for Signal's libsignal in four weeks, work that historically took expert teams years, and it caught two reintroduced bugs.
What you could build with it
A security team at a messaging or fintech company could point an agent-driven formal verification pipeline at its own cryptographic library, getting machine-checked proofs for each release instead of relying on manual audits that take months.
Does it hold up?
Too early to judge: a research framework, not a shipped tool; the verification artifacts are open-sourced for review, but there is no public tool or API yet and the proofs cover one-to-one sessions only, not group messaging or the full app.
Built with AlphaProof Code
- Google DeepMind uses AI agents to verify Signal's core machine codeunite.ai · DeepMind's AlphaProof Code used teams of agents to produce machine-checked Lean proofs for Signal's compiled libsignal code in four weeks, with specifications and verification artifacts released for outside review.
- Google DeepMind melds humans and AI agents to vet Signal's secure messaging techniquestheregister · Agents generated 6.5 million lines of Lean code and flagged only about 1% for human review; the proofs cover the PQXDH handshake and Triple Ratchet for one-to-one sessions.
- DeepMind unveils AlphaProof Code to formally verify Signal protocoltodayforai · DeepMind's blog post introducing AlphaProof Code, a novel agentic framework for formal verification at unprecedented depth, and the Signal verification milestone.
Learn more
First spotted on article: source.
More AI agents
Agent Browser: Vercel Labs' browser automation CLI for AI agentsVercel Labs' open-source CLI that lets AI agents drive a real browser: click, fill forms, and complete web tasks, with…agents · JEV 0.78CowAgent: open-source super AI assistant and agent harness with memory that compoundsAn open-source personal AI assistant and agent harness by zhayujie, back on Hacker News' front page this week: it…agents · JEV 0.78google/skills: Google's own agent skills for its products and Cloudgoogle/skills is Google's official collection of agent skills for Google products and Google Cloud, giving coding…agents · JEV 0.78HorizonSelf-hosted AI news radar in Python: fetches RSS, Hacker News, Reddit, Telegram, X and GitHub concurrently, AI-scores…agents · JEV 0.78anthropics/skills: Anthropic's official collection of 19 Agent SkillsAnthropic's official, curated collection of 19 Agent Skills covering design, documents, coding, and enterprise…agents · JEV 0.77Orca: run a fleet of coding agents side by sideStably AI's open-source desktop orchestrator fans one prompt across 5-30 CLI coding agents, each in its own git…agents · JEV 0.77
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.