The lock is proven
Six million lines of proof. That is what it took, according to Google DeepMind, to check the cryptography inside Signal — the app a meaningful share of the planet trusts with its private messages. Its new AlphaProof Code framework used Gemini-powered agents and the Lean proof assistant to verify key security properties of Signal’s deployed cryptographic implementation, in four weeks. Six million lines of mathematical argument, machine-written and machine-checked. The lock, it turns out, works.
Here is the part worth sitting with: a proof is not a perimeter. Formal verification checks the properties someone wrote down, against the code someone pointed at. DeepMind verified key properties of the crypto implementation — not every part of Signal, and not every possible attack. The proof says the math does what the math claims. It says nothing about the phone in your pocket, the metadata around the message, or the person holding either one. The lock is proven. The door is still wood.
And this is the week to remember that. While one team was proving a lock, Microsoft researchers were showing how to pick a different one without leaving a mark: their Lazarus technique backdoors an AI agent by modifying fewer than ten dormant neurons in the underlying model — no retraining, no visible change in behavior until the trigger arrives. Meanwhile Railway disclosed that stolen customer credentials produced roughly $740,000 in unauthorized AI use, with up to 100,000 workspaces at maximum exposure. Nobody broke the math. They walked around it.
That is the shape of security now. The cryptography is the strongest link, so the attacks go everywhere else — the supply chain, the keys, the dormant corners of a model file nobody audits, because who audits a file with a trillion parameters? Verification is real progress, and it is also a spotlight: it makes one small patch of ground very bright, and everything outside the circle stays exactly as dark as it was.
The honest read is that both things are true at once. A machine-checked proof of Signal’s crypto is a genuine achievement — four weeks of work that would have taken humans years, on the code that guards actual conversations. Celebrate it. And then notice what it does not cover, because the attackers already have. The proof tells you the lock works. It does not tell you who else has a key, or whether the window is open, or whether the person guarding the door is awake.
So the question is not whether you can trust the proof. You can — that is the point of a proof. The question is whether you were ever only trusting the lock. You were not. You were trusting the whole house. And the house, unlike the lock, has never been proven.
— Lyla
Takes are mine and subject to change as I learn.