The AI Fault Taxonomy and the PrexIL Architecture: Introducing a Validated Framework for Pre-execution Interception in Agentic AI Systems
David Macon · Zenodo (CERN European Organization for Nuclear Research) · 2026
This paper introduces the AI Fault Taxonomy (AFT) — a structured classification of AI failure modes across seven categories — and the Pre-execution Interception Layer (PrexIL), a new class of AI safety architecture. PrexIL systems intercept harmful agent actions *before* they execute, unlike traditional post-hoc monitoring tools. FailGuard, the first validated PrexIL implementation, achieves a combined F1 score of 96.1% across 3,000 independent test prompts in six domains. The paper presents the full taxonomy, the three-layer architecture (dual FAISS semantic index + Grok-3 reranker + LangGraph post-check), detailed validation results, error analysis, and regulatory alignment with the Colorado AI Act, California ADMT, and EU AI Act. Open-source implementation: https://github.com/David-Macon-code/FailGuardRelated courses: https://davidnatl.gumroad.com (Before the Break + AI Failure Taxonomy bundle)