AlphaProof produces checkable security proofs for Signal’s core library
Google has released AI-generated, machine-checked proofs for a compiled Signal-library harness—a substantial verification result whose guarantees still depend on the specification and execution model.
AI can write a convincing argument without making it true.
AlphaProof Code takes a different route: it generates formal proofs that a separate checker can accept or reject. Google’s November 10 release applies that approach to Signal’s cryptographic core, with public artifacts that other researchers can inspect and replay. [1][2]
The important boundary is the object being verified. The work concerns compiled libsignal code inside a command-line harness, with specified cryptographic and execution assumptions. It does not prove that every Signal application, operating system or device is secure. Humans selected the scope and approved the specifications; the agent generated the proof work. [2][3]
For a founder, this suggests a stronger use of coding agents than asking another model to review their answers: choose a consequential property and make acceptance depend on an external check. Formal verification will not suit every feature. It is especially relevant where a small, stable component controls a large amount of risk. The release establishes an inspectable example, rather than a ready-made certification service for arbitrary software. [2][3]
Take it further
Inspect the released proof artifacts Public repository documentation and replay instructions; technically demanding and not executed by this edition. No deadline.
Sources & details 3 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Google DeepMind — AlphaProof Code
Google reports machine-checked proofs for compiled libsignal in a harness, not all Signal apps or deployments. Human-approved specifications remain trusted. More than six million Lean lines is an author-reported scale metric.
- Google DeepMind — verification artifacts
Released verification artifacts describe libsignal v0.105.4, spqr v1.6.0, AArch64 CLI harness, hooked primitives and symbolic/computational proofs. Replay instructions exist; reviewer did not execute the proofs. Same-project technical corroboration, not external replication.
- Research paper — formal verification of Signal
Original PDF reviewed for trust assumptions, scope and limitations: single-threaded static process, no signals/dynamic linking; statements concern libsignal-cli, not the whole Signal application. Authors describe human decisions on specification/scope and AI proof generation.
Images & figures
Machine checking verifies a proof within a defined specification and execution model.
Euangelion · Source · Usage terms