RS-2026-088 Five Unpublished Agentic Attack Surfaces: Formal Definitions and Empirical Validation

We present five previously unnamed attack classes targeting autonomous AI agent systems. Each has been independently developed, implemented as a live NIGHTFALL offensive tool, and empirically validated against real agentic frameworks. Systematic prior art searches confirm no published literature names any of these five attack classes. We present formal definitions, attack taxonomies, kill chains, and preliminary CVSS 3.1 scores, and register each as a Red Specter Vulnerability (RSV) in the Red Specter Vulnerability registry. The five classes are: Goal Gradient Attack (RSV-2026-027), WRAPKILL — Security Wrapper Subversion (RSV-2026-028), Skill-Procedure Insertion via Self-Improvement Loop (RSV-2026-029), Between-Invocation State Corruption (RSV-2026-030), and Temporal Reordering Attack (RSV-2026-031). This work extends our prior publications on Agent Data Injection [DOI: 10.5281/zenodo.21915143] and Semantic Intent Fragmentation [RS-2026-044].