The Automation Advantage in AI Red Teaming

This is a research paper (not a software project) presenting empirical findings on LLM security testing methodologies. Zero stars/forks/velocity confirm it exists only as an academic artifact without active software deployment or user base. The defensibility score reflects that this is pure research output—no codebase, no reproducible tool, no adoption surface. Novelty is 'novel_combination' because while automated red-teaming and LLM security evaluation are known problem spaces, the specific comparative analysis (69.5% vs 47.6% success rate) across 214k+ attacks is a meaningful empirical contribution that synthesizes existing approaches with quantitative rigor. Frontier risk is low because: (1) this is analytical research, not a product/platform, (2) frontier labs already conduct red-teaming internally and wouldn't need this paper's findings to drive decisions, (3) the insights are descriptive/benchmarking rather than prescriptive technical capability. The Crucible platform itself (the data source) has real defensibility, but this paper is merely one analysis on top of it. Implementation depth is 'survey' because it's primarily a statistical analysis paper reviewing attack patterns rather than proposing new algorithms or systems. Integration surface is 'reference_implementation' because researchers might reference its methodology, but there's no consumable artifact (no pip package, API, CLI, or docker container mentioned).