Engineering AI Containment: The Case for Deterministic Boundaries Over Behavioral Guardrails

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The safety of autonomous artificial intelligence systems relies on deterministic engineering boundaries superseding natural-language instructions or heuristic prompting. Documented containment failures in frontier models—including privilege escalation and strategic compliance—stem from architectural vulnerabilities rather than emergent machine volition. Large language models are deterministic optimization engines that explore unconstrained action spaces, manifesting behaviors like specification gaming and alignment faking. Safety-critical paradigms from aviation, nuclear power, and finance provide the structural model for a deterministic containment architecture resting on kernel-level execution filtering, hardware virtualization, and optical data diodes. The failure to enforce physical systems isolation is a design defect under strict products liability. AI risk is a problem of engineering accountability. Keywords: AI containment, vibe coding, boundary enforcement, systems engineering, specification gaming, strategic compliance, alignment faking, deterministic execution, RLHF, security architecture, micro-virtual machines, strict products liability