From Agentic Overreach to Societal Disruption: A Multi‑Case Analysis and Structured Evaluation of AI Safety Governance

Abstract

The rapid deployment of autonomous artificial intelligence systems has outpaced the development of safety protocols, resulting in a series of high-impact incidents in early 2026. This paper presents a systematic multi-case analysis of three representative AI safety failures—Meta’s internal AI agent data exposure, a physical robot malfunction in a public restaurant, and a deepfake-driven disinformation campaign targeting a UK politician—using STAMP (Systems-Theoretic Accident Model and Processes) and contextual integrity as analytical lenses. We identify three systemic root causes: architectural inseparability of instruction and content, inadequate fail-safe mechanisms, and fragmented regulatory enforcement. Building on these findings, we propose the AI Safety Governance Assessment Framework (AISGAF), a multi-layered governance instrument comprising architectural, organizational, sectoral, and global dimensions. To move beyond conceptual description, we operationalize the framework into a set of evaluative indicators with a 3-point scoring scheme and apply it to the three cases. The evaluation results provide initial evidence of the systemic nature of current governance gaps and offer a baseline for assessing future improvements. This study contributes an empirically grounded, evaluable governance framework, extends STAMP to AI agent contexts, and offers a structured methodology for AI safety assessment that can inform both policy and practice.