Despite developments in the human aspects of cybersecurity, inclusivity concerns, particularly related to neurodiversity, remain largely unexplored. This paper examines the role that Large Language Models (LLMs) can play in generating code that integrates inclusivity concerns in the development of security systems and services. To do so, a study was conducted using two user stories, each comprising three experimental cases that varied in the level of inclusivity specification from no specification in the first case to a detailed specification in the third During each case, the LLM was tasked to generate a code artifact that implemented the specified requirements. Furthermore, ten repeated code generations were performed in each case to minimize the impact of run-to-run stochastic variation. The generated code artifacts were then subjected to independent evaluations by six different LLMs using a rubric-based assessment. The highest-scoring artifacts based on the rubric-based LLM evaluations were subsequently evaluated by ten inclusivity and ten security experts, and the security assessment was further complemented by static and dynamic analysis using automated tools. The results indicate that greater specification detail is associated with improved inclusivity in LLM-generated code. Statistical comparisons using mixed-effects models revealed no statistically significant differences between human and LLM-based evaluations for inclusivity, whereas a significant difference was observed for security assessments. The findings have implications for research on integrating LLMs within secure software development contexts by highlighting the role of specification details and complementary strengths of human and LLM evaluators for assessing LLM-generated code.
We present RiskOnto, an OWL 2 DL ontology-driven framework for automated cybersecurity risk assessment that integrates MITRE ATT&CK, MITRE D3FEND, MITRE CWE, the NIST Cybersecurity Framework (CSF) 2.0, and NIST SP 800-53 Revision 5. RiskOnto comprises 103,906 RDF triples organized around 121 VulnerabilityProfile archetypes that represent recurring vulnerability patterns and serve as semantic anchors for cross-framework reasoning. Offline materialization of 13 global SWRL rule specifications, organized into three functional sublayers, produced 19,671 reusable semantic assertions (from the 8 inference/explanation rules, R1–R4/E1–E4) and 847 VP-level hasJustification links to typed XAI justification/explanation individuals (from the remaining 5, X1–X5). The core-reasoning sublayer infers NIST CSF impacts and propagates ATT&CK mitigation and MITRE D3FEND recommendations. The explainability sublayer generates rule-grounded assertions covering risk, candidate NIST CSF 2.0 framework-outcome associations, mitigations, and defenses, while the structured XAI sublayer links vulnerability profiles to typed justification and recommendation records that support transparent and auditable risk decisions. Risk is represented using an ordinal Impact × Likelihood model informed by NIST SP 800-30 Rev. 1, assigning scores from 1 to 25 across five criticality tiers, from Very Low to Critical, represented by 352 RiskAssessment individuals. A System Under Test (SUT) integrates with the shared knowledge base through a standardized six-step pipeline using a single OWL import declaration and SUT-specific vulnerability-profile links, without modifying RiskOnto. We validate the framework using WebGoat 2025.4-SNAPSHOT, a deliberately insecure Spring Boot application. Four heterogeneous scanner tools generated 273 raw findings, of which 166 were within the current CWE-representable canonicalization scope and consolidated into 43 canonical security findings; 107 findings remained outside the current CWE-based canonicalization scope. All 43 canonical findings activated at least one VulnerabilityProfile, demonstrating complete workflow coverage of the normalized security findings, while not establishing semantic correctness of every association. The case study demonstrates RiskOnto’s ability to consolidate heterogeneous security evidence, perform reusable cross-framework semantic reasoning, identify framework outcomes potentially associated with vulnerabilities, represent risk, and generate traceable cybersecurity recommendations.
