That isn’t proof that the whole Signal app is secure. The Register says the proof assumes perfect cryptography: it treats the underlying cryptographic operations as ideal, rather than establishing how hard they would be to break in practice. Source
The AI writes proofs. Lean checks them.
DeepMind’s claim reaches beyond an abstract protocol description to machine code, the instructions a processor executes. The Register reports checks of protections including authentication—checking who messages come from—and forward secrecy, which protects earlier messages if keys are later compromised. Source Source
Researchers used models to turn Signal’s security goals into a human-reviewed mathematical specification. Gemini Carbon agents then generated proofs in Lean, a language whose proof checker verifies them against that specification. Whether the proof checks out is not decided by how confident the AI sounds. Source
The agents generated 6.5 million lines of Lean code, then audited it and flagged material for human review, according to The Register. That’s proof code, not 6.5 million new lines bolted onto your messaging app. Source
Four weeks and five people
This wasn’t an AI doing a solo weekend project. The Register reports that a five-person team spent four weeks and “quite a few tokens” on the work. The report gives no itemized cost or like-for-like non-AI benchmark. Four weeks here is not a delivery estimate for every other piece of software. Source
In his announcement, Pushmeet Kohli says the team is open-sourcing its formal specifications and verification artifacts. He names medical-device firmware and power-grid software as possible future applications. Source
Applying the approach beyond cryptographic protocols still requires researchers to decide which protections to demand and how to specify them, DeepMind research scientist Swarat Chaudhuri told The Register. Source
Comments