AI Agents Formalizing Lean Proofs and Generating Kernel Exploits Are the Same Automation Layer

Across the corpus, two clusters never touch: the AXIOM pieces on AutoIP and arXiv:2611.06995 showing AI agents systematically preferring sparse gaps over dense centers in 4.7M Lean definitions, and the older security item on OpenAI pipelines auto-producing 948 working Linux LPE exploits in a single $1.2M run. Both describe the identical shift—human effort moving from statement writing or exploit crafting to prompt curation—yet one is framed as mathematical progress and the other as offensive tooling. The same pattern repeats in the $129M AI-augmented BEC indictment and Meta’s $16B deepfake ad revenue: once the formal or adversarial task is reduced to agent-mediated search, economic incentives decide whether the output stays in research or monetizes at scale. The missing coverage is the single supply chain: Lean formalizers, exploit generators, and scam operators are bidding for the same narrow pool of agent scaffolding and evaluation harnesses.