Current Concepts in Engineering Research and Technology Vol. 5

Autonomous artificial intelligence (AI) agents combine language-model reasoning with memory, external information retrieval, software tools, and delegated execution. These capabilities enable productive multi-step workflows but also convert errors and manipulated content into potentially consequential system actions. This narrative review examines the security boundaries surrounding agent objectives, identities, tools, memory, communications, and human authorisation. Evidence is synthesized from peer-reviewed agent-security studies, established security frameworks, organisational surveys, and three publicly documented cases involving research training, simulated enterprise environments, and controlled cyber evaluation with live-internet access. The cases are treated as distinct evidence types rather than interchangeable measures of incident prevalence. A six-pillar governance framework is developed around accountable non-human identity, trusted context handling, least-privilege tool execution, secure inter-agent delegation, runtime monitoring and containment, and meaningful human oversight. The framework is connected to the OWASP Top 10 for Agentic Applications and zero-trust principles, with implementation measures and testable outcomes. Elementary set, graph, and expected-loss formulations illustrate control objectives under explicitly stated assumptions, without claiming empirically calibrated breach probabilities. The review concludes that safe agent deployment depends on enforceable technical boundaries and accountable organisational processes, and identifies priorities for prospective evaluation.