Harsha Kalutarage

There is a growing use of Artificial Intelligence (AI) systems in high-stakes environments such as healthcare and transportation, where errors or misuse can lead to serious and potentially irreversible consequences. Existing regulatory frameworks remain fragmented and often inadequate to address the complexity of AI development in these settings. Although ethical principles such as safety, fairness, accountability, transparency and human oversight are widely recognised, their translation into enforceable legal obligations remains uneven across jurisdictions and sectors. The central question addressed in this paper is how AI use in high-stakes environments should be regulated to achieve effective and enforceable obligations. The paper makes four contributions. First, it develops a three-tier Ethics Pyramid, a framework that distinguishes foundational ethical values, second-level governance principles, and third-level operational mechanisms. The framework functions as an allocation model, helping regulators identify which ethical principles should be juridified and at what level of governance this should occur. Second, the paper evaluates regulatory theories, strategies and instruments and argues that a polycentric, multi-layered legal framework is better suited to high-stakes AI governance. Third, it conducts a comparative analysis of high-stakes AI regulation in the UK, the EU, the US and China, alongside a sectoral examination of healthcare, autonomous vehicles, criminal justice and public sector decision-making. The analysis reveals significant variation in regulatory maturity. Fourth, it develops a polycentric, multi-layered legal architecture illustrated through case studies of AI use in UK prisons and autonomous vehicles, highlighting the contrast between fragmented and mature regulatory regimes.

Insider threats, originating from current or former employees, contractors and trusted business partners who are with authorised access to organisational resources continue to pose a critical security challenge across multiple sectors. These threats may be intentional (malicious) or unintentional (accidental) and are shaped by the interaction between technical and human factors. This systematic literature review analyses 121 studies published between 2014 and October 2025 to address two key research questions: (1) which human factors contribute to insider threats, and (2) whether existing technical methods, particularly machine learning approaches, have been applied to predict or detect insider threats on the basis of human factors. The review introduces a taxonomy of human factors influencing insider behaviour and provides an examination of technical approaches, including detection methods. The findings highlight a substantial gap in machine learning research directly linking human factors to insider threat detection, largely due to the absence of datasets that integrate behavioural and contextual human attributes. By synthesising current knowledge and identifying these limitations, this study outlines key challenges and research opportunities, underscoring the need for predictive models that capture the complex role of human factors in insider threats especially in unintentional insider threats.

Embedded systems, including the Internet of things (IoT), play a crucial role in the functioning of critical infrastructure. However, these devices face significant challenges such as memory footprint, technical challenges, privacy concerns, performance trade-offs and vulnerability to cyber-attacks. One approach to address these concerns is minimising computational overhead and adopting lightweight intrusion detection techniques. In this study, we propose a highly efficient model called optimized common features selection and deep-autoencoder (OCFSDA) for lightweight intrusion detection in IoT environments. The proposed OCFSDA model incorporates feature selection, data compression, pruning, and deparameterization. We deployed the model on a Raspberry Pi4 using the TFLite interpreter by leveraging optimisation and inferencing with semi-supervised learning. Using the MQTT-IoT-IDS2020 and CIC-IDS2017 datasets, our experimental results demonstrate a remarkable reduction in the computation cost in terms of time and memory use. Notably, the model achieved an overall average accuracies of 99% and 97%, along with comparable performance on other important metrics such as precision, recall, and F1-score. Moreover, the model accomplished the classification tasks within 0.30 and 0.12 s using only 2KB of memory.