Can you trust the maths behind your messages? Google DeepMind just gave the world a serious reason to say yes. The lab's AlphaProof Code — a multi-agent framework that turns Gemini-powered agents loose on formal verification — has produced a machine-checked proof of key security properties in the deployed implementation of Signal's cryptographic protocol. It took four weeks and roughly six million lines of Lean.
What AlphaProof Code actually did
AlphaProof Code splits a hard verification job across a team of agents and lets them work in Lean, the proof assistant that checks every step of a mathematical argument. DeepMind's announcement, published on November 10, describes a multi-LLM agentic framework: Gemini-powered agents generated the proofs and checked them, producing the first formal verification of a widely deployed cryptographic protocol by AI. Four weeks of work, six million lines of proof, and an argument a machine can check but no human could read through unaided.
Why Signal was the right test
Signal isn't a demo. Its protocol guards the messages of millions of people, and the verification targeted the deployed implementation rather than a simplified model. That choice is what gives the result its bite: it says something concrete about the code people actually run, not about an ideal sketch of it. Signal's protocol was a natural target precisely because so much rides on it, and the story reached the front page of Hacker News on November 10.
Tests can miss things. A formal proof leaves no step unexamined: either the argument holds, or the proof assistant says no. That's why cryptographic protocols are such a natural target — and why this result matters more than a benchmark win.
What four weeks means for security work
Proofs at this scale once meant years of specialist labour. If AI agents can produce them in a month, the cost of auditing the software that guards our secrets drops fast. It changes who can afford to check the crypto under the hood — and how often they do it. DeepMind has also open-sourced the proof artifacts and published a technical paper, so others can pick the work apart for themselves.
DeepMind has been pushing AI further into security research all year — the lab's agents also found two weaknesses in 1024-bit RSA, and Google is building new TPU chips built for agentic AI. Formal verification is the other side of that coin: not just finding flaws, but proving the good parts really hold.
Signal's users won't notice anything different on their phones. That's the point — the math underneath just got a lot harder to doubt.
Originally published on Rohan Tech Official.
Top comments (0)