From taxonomy to stability: a CISO-ready engineer-deep architecture and best-practice blueprint for bounded autonomy in agentic cyber defense (AI vs AI)

Purpose This paper aims to address “escalation instability” in Agentic Cyber Defense by proposing a framework for bounded autonomy that balances machine-speed response with operational safety. Design/methodology/approach The study develops a blueprint for the Bounded Autonomy Controller (BAC) acting as a runtime gate between reasoning and execution, introducing an Escalation Ladder (L0–L3) to categorize action authority and mandating a structured Decision Packet for deterministic authorization. Findings Traditional taxonomies are insufficient for autonomous systems because they fail to capture action authority. Through qualitative case study validation, this study demonstrates that bounded autonomy mitigates prompt injection risks and prevents “false ransomware” escalation spirals that otherwise cause business-stopping disruption. Research limitations/implications The study highlights that evaluating autonomous agents requires moving beyond static data sets toward dynamic adversarial “agent-vs-agent” simulations. Practical implications A four-phase rollout plan is proposed to operationalize NIST AI RMF principles into technical guardrails, enabling Security Operations Centers (SOCs) to transition safely to automation. Social implications The framework protects digital availability and critical business processes by engineering trustworthy, predictable systems that prevent unintended autonomous “self-damage” during cyber conflicts. Originality/value This paper provides a unique CISO-ready, engineer-deep blueprint for trustworthy Autonomous Cyber Defense systems, introducing novel stability-centric metrics such as the Self-Damage Index (SDI) and False Escalation Rate (FER) to quantify autonomous performance.