Source time Edition date
Introducing AlphaProof Code: a breakthrough in verifying complex software at depth
推出 AlphaProof Code：深度验证复杂软件的重大突破
AlphaProof Code matters to security engineers, but Google DeepMind's three-million-line Lean proof effort for Signal's core encryption shows formal verification still demands massive agent labor before it scales.
Judgment
Signal’s security leadership can now weigh AI-assisted formal verification for protocol releases, but the four-week, over-six-million-line Lean effort still requires expert specifications and human audit of one percent of proofs, so it should be reserved for high-value cryptographic changes rather than routine code review.
Basis
Ehren Kret, Signal’s CTO, called the use of AI agents to translate live code into valid Lean theorems “a major breakthrough” and said rigorous verification gives Signal greater confidence in its privacy guarantees. The mechanism is not a single model: Google DeepMind’s AlphaProof Code combines human experts, who review every formal specification, with Gemini Carbon agents that generate over six million lines of Lean; Lean then machine-checks the proofs. Agents flag just one percent of the codebase for human audit, and the whole process took just four weeks. To test bug detection, the team reintroduced two previously fixed bugs into an isolated Signal codebase, and the system produced machine-checked proofs that both buggy versions were insecure. DeepMind says this is the first formal verification of a widely deployed cryptographic protocol implementation at this depth.
Unclear The source does not say how often the agents’ audit triage misses a flawed proof, nor whether the one percent human-review slice was chosen without seeing the proof’s conclusions.
Next Check whether the open-sourced Lean artifacts include the two reintroduced-bug proofs and a reproducible command that flags the insecure versions.