Resilience of AI Systems: Detection, Containment, Recovery, and Re-entry

AI systems will fail. The central engineering question is what remains controllable once they do.

Resilience of AI Systems develops a rigorous framework for reasoning about AI after failure: when an incident first becomes detectable, how long detection takes, how containment limits the blast radius, what can and cannot be rolled back, how degraded operation should be structured, and what evidence is required before unrestricted operation can safely resume. Rather than treating recovery as a single metric, the book decomposes incident damage into distinct stages and shows which parts are determined by instrumentation, which are affected by response policy, and which remain outside the reach of any post-incident intervention.

Across detection theory, graph-based containment, hysteresis and review queues, degraded-safe modes, rollback and compensation, redundancy, belief-based re-entry, incident reconstruction, and assurance, the text connects mathematical structure to concrete operational decisions. A running deployment case, explicit provenance labels, boundary statements, CPU-only computational labs, and proof-oriented exercises make the book suitable for graduate study as well as technical design and review.

For students, researchers, safety engineers, and system architects, this volume provides a unified way to ask not merely whether an AI system can recover, but what recovery can actually guarantee.