Reproduce “Safety Is $Π_1$, Provability Is $Σ_1$: One Framework for the Hugging Face Breach and OpenAI's Mathematical Proofs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jinho Kang, Jinsook Kim

The paper compares AI-assisted security testing and theorem proving through a proposer–verifier loop: a model offers candidate exploits or proofs, and a checker evaluates each one. It formalizes witnesses as finite evidence and shows that, for undecidable witness classes, any bounded computable search can miss infinitely many existing witnesses. This limits the result to stated assumptions; checking candidates cannot establish universal safety.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.