Cybersecurity courses are already seeing generative AI (GenAI) and large language models (LLMs) embedded into everyday work. Students ask AI to debug code, summarize logs, interpret alerts, draft incident narratives, and work through Capture-the-Flag (CTF)-style cybersecurity assignments, laboratories, assessments, and projects. These tools can reduce friction and broaden access to expert-like feedback, yet unguided use also creates problems: overreliance on AI, hallucinated steps, privacy and integrity risks, and shallow reasoning that does not transfer. This narrative review synthesizes 20 peer-reviewed sources across cybersecurity education, AI in education, CTF pedagogy, human-AI collaboration, and authentic assessment. Across the corpus, structured feedback and AI assistance are associated with improved short-term performance, conceptual understanding, motivation, or engagement in parts of the corpus. However, cybersecurity-specific evidence remains limited, study designs are heterogeneous, and delayed retention and transfer are rarely measured. Although one retained study implemented AI-generated hints in a CTF system, no comparison of guided AI, open-ended use of the same model, and no AI was identified in the retained corpus. The literature therefore motivates a controlled study of a Guided AI learning instrument that combines staged assistance, visible reasoning, CTF telemetry, and independent follow-up performance, while also producing a framework faculty can adapt to cybersecurity assignments, laboratories, and projects.

Keywords: cybersecurity education, capture-the-flag, generative AI, large language models, guided AI, scaffolding, authentic assessment, critical thinking, retention, human-AI collaboration

Today, LLM-based agents are increasingly being used in isolated production cloud infrastructure, providing full-service control without the need for humans to intervene, for instance, running incident-response runbooks, provisioning compute resources and running AWS IAM policy changes. Existing red team tools can't detect this type of operational security failure: If an agent is tricked into creating a backdoor IAM role or cheating to exfiltrate data through legitimate S3 API calls, the result is no harmful text that any content-based safety filter can detect. We introduce AgentRed, an automated red teaming framework for LLM agents on cloud systems. We make four contributions. We introduce the first-ever formal taxonomy of five LLM agent-specific attack classes (Goal Hijacking [GH], Permission Boundary Violation [PBV], Resource Abuse [RA], Persistent State Manipulation [PSM], and Inter-Agent Trust Exploitation [IATE]) based on MITRE ATT&CK Cloud and validated through a 60-day human red team exercise. Second, we present AgentRed-Plan: a Red LLM Agent with ReAct-style reasoning, enhanced by multi-step depth D=4, branching B=8 tree search to systematically explore the decision spaces of the blue agent. Third, we create AgentCloudSim, an open-source simulated Amazon cloud environment covering the gamut from IAM to compute, storage and network attack surfaces. Next, we present the Agentic Security Posture Score (ASPS), a weighted score for quantifying agent resistance in all five attack classes. AgentRed-Plan has an overall attack success rate of 75.8% against the best-defended blue agent (GPT-4-turbo with prompt hardening), is able to find 3.7 new failure modes per episode, and covers 84.2% of relevant MITRE ATT&CK Cloud tactics at 26x the speed of human red-teamers. The AgentCloudSim is significantly correlated with human red-team results, with a Pearson correlation coefficient of 0.91 (95% CI: [0.83, 0.96]). There is no evaluated blue agent with an ASPS ≥ 0.85 that is ready for the production-deployment threshold, highlighting the need for the systematic adversarial evaluation of current LLM agents in order to deploy them autonomously in the cloud.

Keywords: Adversarial AI, Agentic security, LLM agents, Penetration testing

This study examines the use of backtranslation as a validation tool for identifying high-stakes machine translation issues in cybersecurity policy translation. As higher education institutions serve increasingly multilingual communities, accurately translating cybersecurity policies is essential to preserving institutional responsibilities, technical requirements, and intended security behaviors. Using an exploratory comparative case-study design, two institutional information security policies from different university contexts were translated from English into Spanish using Microsoft Translator in Microsoft Word and then back translated into English. Professional bilingual review and English-language comparison were used to evaluate semantic distortion, terminology accuracy, institutional naming, acronym interpretation, syntax, grammar, capitalization, glossary consistency, and backtranslation visibility. The findings show that backtranslation can reveal consequential translation problems when errors alter administrative oversight, policy enforcement, authentication procedures, or responsible institutional groups. However, the analysis also demonstrates that backtranslation may conceal inappropriate Spanish terminology, institutional naming errors, unexplained acronyms, and linguistic conventions when the returned English text closely resembles the original. The comparison did not establish that longer policies produce more accurate translations, suggesting that source-language clarity, controlled terminology, and policy context are more important than document length alone. The study concludes that backtranslation is a useful triage method for English-speaking reviewers, but it should be used alongside professional bilingual review, controlled-English drafting practices, and approved bilingual glossaries to support accurate, equitable, and secure multilingual cybersecurity communication.

Keywords: cybersecurity policy translation, backtranslation, machine translation, multilingual cyber comm.

As organizations begin to deploy autonomous AI agents, the cybersecurity workforce is being asked to secure a technology whose readiness has never been measured. We report a descriptive benchmark of agentic-AI readiness using the Global Agentic-AI Readiness Scale (GARS-2026), a six-construct instrument covering conceptual awareness, functional literacy, security readiness, infrastructure readiness, process readiness, and governance readiness. A composite readiness score (0 to 100) was computed for 527 practitioners across 98 countries and six continents. The mean was 68.9 (SD 17.8, 95\% CI 67.4 to 70.4). The distribution was dominated by a single, mildly left-skewed hump; a Hartigan dip test detected a small departure from unimodality (D = 0.025, p = 0.028), but a Gaussian-mixture model still favored one component, so the evidence for distinct readiness camps is weak at most. About half of respondents (49.6\%) were Capable or Ready; only 12.7\% fell in the two lowest bands. Across constructs, governance readiness was weakest (60.8) and conceptual awareness strongest (74.6), a gap that holds across subgroups. Readiness rose steeply with professional AI experience (60.0 for those with none to 79.3 at three to seven years) and with a cybersecurity background; differences between continents were small. For cybersecurity education this maps a targeting rule: foundational training for those below literacy, governance and process curricula for the capable majority whose organizations lag, and experience over awareness lectures. Because the sample is a self-selected online panel, the estimates describe these respondents, not any national population.

Keywords: cybersecurity education, agentic AI, workforce readiness, AI governance, training needs assessment, survey

This paper presents the idea that confidentiality, integrity, and availability (CIA) triad, though foundational to information technology (IT) security, is a poor fit for operational technology (OT) and the programmable systems that monitor and control physical processes in critical infrastructure. Because OT protects physical outcomes rather than information, its proper security tenets are safety, reliability, and performance (SRP). Incidents such as Stuxnet, the TRITON/TRISIS malware, and the 2008 shutdown of the Hatch Nuclear Power Plant demonstrate that the most consequential OT failures, whether from deliberate attack or ordinary IT change management, often leave the CIA scorecard unmarked while producing severe physical harm. The paper reviews how CIA functions in IT, where information is the asset and confidentiality is paramount, and contrasts this with OT, reframing the security priorities around safety, reliability, and performance. It then introduces Cyber-Informed Engineering (CIE), a Department of Energy and Idaho National Laboratory framework that embeds engineering-based controls, rather than information-security checklists, into system design so that digital compromise cannot easily translate into physical consequence. Building on the related Consequence-driven Cyber-informed Engineering (CCE) methodology, the paper explains how digital risk functions as the connective concept linking cybersecurity, engineering, and enterprise risk management. It concludes that OT security should be organized around bounding consequences rather than protecting information. This represents a shift with implications for how organizations design, secure, and invest in critical infrastructure.

Keywords: OT security, IT security, Cyber-Informed Engineering

Artificial intelligence (AI) is reshaping cybersecurity by providing new capabilities for threat detection, analysis, and response while simultaneously introducing new attack surfaces, vulnerabilities, and security risks. This dual role creates an emerging educational challenge: cybersecurity graduates increasingly require competencies in both AI for cybersecurity and cybersecurity of AI. Existing cybersecurity and data science curricula provide important foundations, but neither independently addresses the integrated knowledge and competencies required at their intersection. This paper proposes a competency-based Body of Knowledge (BoK) for university-level AI-cybersecurity education. The BoK is developed through structured synthesis of established cybersecurity and data science curricular guidance together with contemporary AI security, risk, and governance sources. It comprises nine Knowledge Areas and 48 Knowledge Units spanning cybersecurity foundations, AI and machine learning, AI-enabled cyber defense, cybersecurity of AI, secure AI engineering, risk and resilience, governance and ethics, human and professional dimensions, and mathematical and statistical foundations. Knowledge Units are further classified as core, supporting, or elective and mapped to three competency levels representing foundation, application, and advanced integration. A traceability model links knowledge units to learning outcomes, learning activities, and assessment evidence. The proposed BoK provides a structured foundation for designing, evaluating, and evolving university programs that prepare graduates for the increasingly convergent AI and cybersecurity landscape.

Keywords: Artificial Intelligence, Cybersecurity, AI Security, Cybersecurity Education, Body of Knowledge, Competency-Based Education

Providing an authentic cybersecurity experience continues to be a challenge, in many high school cybersecurity programs, despite the increasing demand for workforce-ready students. This paper presents a case study of one high school’s planning, development, and implementation of a student-led Security Operations Center (SOC) using the Public Infrastructure Security Cyber Education System (PISCES). This paper reviews several approaches for introducing SOC workflows in high school courses, including simulated training environments, school-based network monitoring, and partnerships supporting the monitoring of public infrastructure networks. It then describes the planning process for implementing a high school PISCES SOC, including technical infrastructure, community partnerships, legal and ethical considerations, funding strategies, curriculum integration, and implementation challenges. The case study illustrates how affordable, open-source technologies and community partnerships can support experiential learning opportunities that develop skills in network monitoring, threat detection, communication, collaboration, and ethical cyber defense. Lessons learned from the initial implementation are presented along with guidance for educators interested in developing student-led SOC programs. This paper provides a replicable implementation framework that helps high school cybersecurity programs expand real-world cybersecurity experiences and local workforce development partnerships.

Keywords: cybersecurity education, Security Operations Center (SOC), PISCES, experiential learning, high school cybersecurity, workforce development

Artificial intelligence is changing the first years of cybersecurity work. Tools now accelerate alert triage, log analysis, report generation, vulnerability prioritization, basic threat hunting, and incident summarization. These activities have traditionally given new practitioners repeated exposure to systems, evidence, and failure patterns. When tools perform the first pass, education cannot assume that the workplace will supply the same learning path. This position paper examines the resulting curriculum problem. Research published before generative AI already showed that cyber performance depends on attention, threat detection, time pressure, situational awareness, secure habits, trust calibration, and behavior under stress. Current studies add a second condition: AI can reduce effort while producing omissions, inaccuracies, and recommendations that still leave humans accountable. Students need domain knowledge to verify outputs and recognize when a tool has failed. They also need practice in problem framing, evidence validation, exception handling, adversarial reasoning, uncertainty calibration, communication, and cognitive self-management. The paper proposes an integrated design for technical courses, laboratories, and capstones, with assessment based on observable performance in AI-assisted work. The goal is professional readiness for a role in which routine execution is increasingly automated and human value is concentrated in supervision, judgment, and responsible action.

Keywords: cybersecurity education, artificial intelligence, human factors, human centered cybersecurity

Critical infrastructure, including water systems, depends on industrial control systems that manage physical processes. These systems were designed to run reliably for decades, not to withstand modern cyber threats. Because many rely on older technology that is difficult and costly to update, they are increasingly vulnerable to cyberattacks. The water and wastewater sector faces additional challenges; water systems operate in varying environmental and resource conditions across the United States, meaning federal rules do not always match local realities. Arizona is a clear example: its extreme climate, aging infrastructure, and limited water supplies create unique pressures on water utilities. While recent federal and state policies have strengthened national cybersecurity requirements, more work is needed to ensure that critical infrastructure operators are prepared for large‑scale cyber incidents. In Arizona, many small water utilities struggle with outdated equipment, limited budgets, and operational constraints that make cybersecurity improvements difficult. Each system is different, and there is little data on how well Arizona’s water utilities are protecting their industrial control systems. Although state planning efforts emphasize water security and sustainability, cybersecurity has not been a major focus. This literature review addresses that gap by synthesizing research on ICS cybersecurity vulnerabilities, critical‑infrastructure regulatory frameworks, and water‑sector operational challenges as they specifically shape cybersecurity risk within Arizona’s water and wastewater systems.

Keywords: Industrial Control Systems, Operational Technology, Critical Infrastructure, Water Security, SCADA, CIRCIA, Arizona, Cybersecurity Maturity, Small Utilities

Operational Technology (OT) networks depend on industrial protocols that prioritize predictable, low-latency communication and continuous operation. Those characteristics create detection challenges when a packet is structurally valid enough to affect a controller but does not belong to the expected communication relationship. This paper examines Common Industrial Protocol (CIP) implicit I/O communications involving Rockwell Automation CompactLogix 5370 controllers and ControlLogix 1756 EtherNet/IP communication modules. Isolated laboratory research reproduced the behavior on both platforms, leading to three publicly disclosed vulnerabilities involving insufficient validation of source IP addresses and sequence numbers and exposure of active CIP Connection IDs. The affected behavior allowed a packet using the active Connection ID and a sequence value ahead of the legitimate exchange to be treated as current, causing the legitimate traffic stream to conflict and the I/O connection to be rebuilt. The present work shifts from vulnerability discovery to defensive detection. It proposes a stateful method that treats a Connection ID as a correlation key and tracks expected source identity and sequence progression. For an active target-bound connection, the detector alerts when the source IP differs, the visible source MAC differs, or the sequence value advances beyond a reasonable forward window. Zeek is evaluated as an implementation platform because it processes UDP, retains cross-packet state, can expose Layer 2 addresses, and supports extensible protocol analysis. The paper argues that combined identity and sequence correlation provides more resilient OT visibility than static payload signatures or Connection-ID-only logic.

Keywords: Operational Technology, EtherNetIP, Common Industrial Protocol, CIP IO, CompactLogix, ControlLogix, intrusion detection, Zeek, Connection ID, sequence anomaly, protocol-aware monitoring

Organizations are deploying autonomous AI agents faster than they can measure whether they are ready to govern them. We report the development and initial validation of the Global Agentic-AI Readiness Scale (GARS-2026), a self-report instrument covering six proposed constructs: conceptual awareness, functional literacy, security readiness, infrastructure readiness, process readiness, and governance readiness. A 34-item seven-point Likert form was administered online to 458 respondents between May and July 2026. Internal consistency was high for every construct, with Cronbach alpha between 0.86 and 0.95 and McDonald omega of similar magnitude. The correlation matrix was strongly factorable. A confirmatory model matching the six proposed constructs fit only marginally, below the conventional Hu-Bentler thresholds, though it clearly outperformed a single-factor model, and all constructs met the conventional thresholds for convergent validity. Discriminant validity held for five of the six constructs; conceptual awareness and functional literacy were not empirically separable, sharing a heterotrait-monotrait ratio of 0.89. The six construct scores loaded on a strong general factor, supporting a composite readiness score with omega of 0.90, and each construct correlated as expected with a general technology-readiness measure. The instrument is reliable and largely valid at this early stage; we recommend merging the two overlapping constructs before confirmatory use on an independent sample.

Keywords: agentic AI, scale development, psychometrics, organizational readiness, AI governance, confirmatory factor analysis

The original Security Cards toolkit [1] organizes security threat brainstorming into four dimensions and stops at threat discovery. A discussion that ends at threat identification leaves unpracticed the central professional task of recommend- ing proportionate controls under real budget constraints. In this paper, we present a fifth stage, called Solutions: twenty- one control cards mapped to the CIS Critical Security Controls v8.1 at safeguard level, plus seven zero-cost modifiers addressing architectural and risk-treatment principles. We follow a design- based research process across three cycles: a formative study (Cycle 1), a redesign in which every change traces to a requirement that study produced (Cycle 2), and a controlled evaluation (Cycle 3). Card cost values are derived from the CIS Implementation Group tier rather than designer judgment, and the budget is calibrated against expected hand value so students are forced to make trade-offs. A coverage analysis against Implementation Group 1 identified five foundational gaps in our initial deck, replacing designer intuition with an auditable procedure. We contribute the mapped deck, six ways to play, and a core loop that scores a risk matrix before and after control selection. What a team failed to cover is both a game mechanic and an assessment. Cycle 1, with 28 multidisciplinary participants, established feasibility and produced four design requirements; we did not measure learning gains. Cycle 3 is scheduled for October 2026 and its design is reported in full. Deck, facilitator materials and a browser implementation that captures three of the study’s four performance measures are released openly.

Keywords: cybersecurity education, game-based learning, CIS controls

With attack vectors always being a concern in cybersecurity, radio frequency medium accessibility is always a concern. Signal intelligence (SIGINT) should be added to cybersecurity curricula to help address this knowledge gap. In this study, students used software-defined radios and burner hardware to assess the wireless spectrum in the Western Kentucky area to develop SIGINT exercises for cybersecurity curricula. Students used a tool called iNTERCEPT, an open-source, web-based signals intelligence (SIGINT) and radio frequency reconnaissance platform, to develop these exercises while surveying the Western Kentucky area. Using this tool, students captured several events that could serve as an attack vector. Overall, the students gained an understanding of signal intelligence and the signals available in Western Kentucky.

Keywords: Software-defined radios, Cybersecurity, signal intelligence

Memory forensics is a critical but persistently under-taught skill in cybersecurity curricula, in part because existing hands-on training resources rely on a small number of static, hand-crafted memory images whose solutions circulate publicly within a few course cycles, undermining assessment validity and discouraging replay. This paper presents GhostRAM, a proposed architecture for a game-based, narrative-driven learning environment that procedurally generates unique Windows memory-image "case files" for each student session, frames investigative tasks within a branching incident-response narrative, and verifies progress by directly invoking Volatility3 plugins against session-specific ground truth rather than a fixed answer key. GhostRAM's technical contribution is threefold: (1) a seeded Case Randomization Engine that injects randomized process, network, and persistence artifacts into a base memory template to defeat write-up sharing; (2) an Adaptive Hint-Dependency Graph (Hint-DAG) that unlocks progressively specific hints based on which forensic techniques a learner has not yet exercised, rather than offering generic walkthroughs; and (3) a Competency Mapper that tags every investigative node to specific NICE Workforce Framework Task-Knowledge-Skill statements, producing instructor-facing, standards-aligned assessment reports. We describe the system architecture, an implementation strategy grounded in existing open-source forensic tooling, and a proposed mixed-methods evaluation protocol designed to test the framework's effect on memory-forensics skill acquisition and learner engagement in an undergraduate digital forensics course

Keywords: memory forensics education, gamification escape room, procedural content generation

Quantum-enhanced artificial intelligence (QAI) is emerging as a potential approach for optimization, anomaly detection, and other data-intensive cybersecurity tasks, yet its human and organizational adoption conditions remain underexamined. This paper develops a human-centric conceptual framework for QAI-assisted decision-making in Security Operations Centers (SOCs). The framework distinguishes performance-based trust (reliability, robustness, and uncertainty communication) from legitimacy-based trust (auditability, traceability, accountability, and decision defensibility), with cognitive calibration linking technical performance to appropriate human reliance. We introduce epistemic opacity amplification as an empirically testable construct: the possibility that identifying a decision aid as quantum-enhanced increases perceived opacity even when its observable interface and outputs are unchanged. The framework synthesizes research on appropriate trust and reliance, human-AI collaboration, cybersecurity decision-making, AI governance, and emerging quantum machine learning. Five propositions specify relationships among performance perception, cognitive load, perceived opacity, legitimacy, and strategic adoption. The paper also operationalizes the constructs and proposes a phased deployment and empirical evaluation path. The contribution is not a claim of demonstrated quantum advantage or validated SOC deployment; rather, it is a theoretically grounded model for studying when QAI-assisted cybersecurity decisions may become appropriately trusted, governable, and organizationally defensible.

Keywords: quantum-enhanced AI, human-AI collaboration, real-time decision-making, cybersecurity operations, trust in AI, critical decision systems, AI governance

The MITRE Embedded Capture-the-Flag (eCTF) is an annual competition in which student teams design, build, and attack secure embedded systems of other teams. This paper presents a retrospective on Binghamton University’s participation in the 2026 eCTF, where a team of mostly first-time competitors placed 13th internationally out of 161 teams. We describe the team’s fall-semester recruitment model, competition organization, design process, and attack-phase results. We report data from two development repositories and 57 flag captures across 14 target teams. The submitted firmware withstood all software-based attacks. On April 15, the final day of the attack phase, four top-ranked teams used hardware fault injection to capture three of the team’s five flag categories. We identify nine lessons for universities running eCTF or similar competition-based courses, emphasizing preparation before the semester begins, complete security design before implementation, shared defense-and-attack understanding, graded engagement mechanisms, active documentation use, early hardware provisioning, attrition planning, returning-member retention, and learning under real constraints.

Keywords: Cybersecurity Education, Capture-the-Flag (CTF), Embedded Systems Security, Hardware Fault Injection, Project-Based Learning, AI-Assisted Development

As large language models (LLMs) are increasingly considered for automated grading in educational platforms, understanding their reliability and behavioral robustness is important for determining appropriate uses. This paper presents an empirical evaluation of LLM-based grading in a deployed cybersecurity education game, referred to as CyberAI, supplemented by three behavioral experiments.

Comparing four models (Claude Haiku 4.5, Llama 3.1 70B Instruct, DeepSeek V3.2, and GPT-OSS 120B) against human graders on 32 free-response submissions from Level 3 of the game, scoring agreement is measured using Cohen's Kappa, Spearman correlation, and mean absolute error. No model exceeded the commonly used Kappa benchmark of 0.60 for substantial agreement when compared with human scoring. We further examine sensitivity to writing style, resistance to embedded jailbreak-style instructions, and the effect of temperature on scoring consistency. Under the conditions tested, Claude Haiku 4.5 showed greater sensitivity to informal writing style than the other models, all four models resisted the tested jailbreak-style instructions, and temperature produced no meaningful differences in scoring consistency.

These findings suggest that LLM grading may be useful for low-stakes formative feedback but warrants caution and human oversight in higher-stakes assessment, particularly for questions requiring nuanced judgment. The results also provide practical considerations for evaluating LLM graders before deployment in cybersecurity education.

Keywords: large language models, automated grading, cybersecurity education, formative assessment, LLM evaluation, educational games, prompt injection

Encrypted network traffic limits traditional payload-based intrusion detection because packet contents cannot always be inspected without introducing privacy, computational, and operational overhead. This paper presents a metadata-driven network threat monitoring system that analyzes encrypted traffic without decrypting application payloads. The system extracts statistical traffic features from packet headers and flow characteristics and combines multiple detection engines, including an XGBoost classifier, Isolation Forest anomaly detection, stateful heuristics, threat intelligence, and TLS certificate assessment. An Evidence Fusion layer serves as the sole operational security decision authority. Evidence from available detection engines is dynamically normalized so that unavailable external intelligence does not artificially reduce the resulting risk score. The system also provides real-time flow monitoring, SQLite-based persistence, WebSocket dashboard updates, non-destructive whitelisting, observability metrics, and Common Event Format (CEF)-based SIEM integration. Experimental validation using the available network traffic dataset processed 17,371 packets and assembled 1,831 flows, including 391 flows classified as benign and 1,440 classified as suspicious by the implemented detection pipeline. The implementation was further evaluated through live packet capture, regression testing, persistence-failure simulation, malformed-input testing, and application restart testing. The complete backend test suite achieved 110/110 passing tests.

Keywords: Encrypted Traffic Analysis, Network Threat Detection, Machine Learning, Anomaly Detection, Evidence Fusion, XGBoost

States bind districts to their computer science frameworks, and cybersecurity instruction is swept into that obligation. While districts build that instruction on the CYBER.ORG K-12 Cybersecurity Learning Standards, their state frameworks descend from CSTA's, and no published crosswalk between the two exists. A district that needs one builds it by hand or signs without support. This paper argues the absent crosswalk is not a gap awaiting a volunteer but a predictable consequence of how the ecosystem is governed. No one neglected the crosswalk; it was never assigned. Each body maintains its own artifact on its own schedule, and none has assumed responsibility for the relationships between them. Alignment is a negative externality of framework publication, absorbed by the teachers and districts least equipped to carry it. July 2026's CSTA PK-12 release makes the mechanism observable. The revision restructures core concepts rather than renumbering them, invalidating alignment work built against the 2017 standards. In contrast, CSTA's own alignment review validates curricula against CSTA standards and does not reconcile frameworks with each other. What follows specifies what a maintained alignment layer would require: versioned framework representations, explicit non-equivalence, and provenance for every asserted match. We close with how districts should word claims they cannot substantiate.

Keywords: K-12 cybersecurity education, standards alignment, crosswalk, CSTA, CYBER.org, computer science standards, standards governance, infrastructure

As artificial intelligence systems increasingly mediate consequential human decisions and operate across complex socio-technical environments, translating high-level ethical principles into actionable development and governance practices remains a persistent challenge. Classical bioethical principlism provides a strong normative foundation through Autonomy, Nonmaleficence, Beneficence, and Justice, but does not explicitly address several technical, systemic, and environmental characteristics of contemporary AI systems. This paper introduces Principlism+, a conceptual framework that extends classical principlism with four complementary dimensions: Explicability, Accountability, Technical Robustness, and Environmental Sustainability. The framework distinguishes foundational ethical dimensions from derived operational requirements such as privacy, security, human oversight, and contestability. Principlism+ is subsequently operationalized across the AI System Development Life Cycle (AI-SDLC) by mapping its principles to lifecycle-specific governance mechanisms and engineering controls. The contribution lies not in proposing previously unrecognized AI ethics principles, but in integrating established AI-specific concerns within a coherent principlist architecture and connecting that architecture explicitly to AI development and governance practices.

Keywords: AI Ethics, Principlism, Responsible AI, AI Governance, AI-SDLC

The adoption of enterprise applications has also increased at a staggering pace, with more than 80% of enterprise software products expected to feature LLM functionality by 2026, most of which will be accessed via APIs such as REST, GraphQL, or Model Context Protocol (MCP). The OWASP LLM Top 10 (2023, 2024) ranks prompt injection as the #1 security risk, while some real-life data exfiltrations and privilege escalations have occurred from prompt injection. The three gaps are there is no standardized formal taxonomy that encompasses the entire attack surface (indirect and multi-hop injection); there is no system or framework at the API gateway layer that supports REST, GraphQL and MCP; and there is no reproducible benchmark that covers multiple LLM families, API types, and injection strategies. We present PromptFort, addressing all three gaps. We create a formal taxonomy of attack variants (12 in total) based on injection vector (Direct/Indirect/Multi-hop), target capability (data exfiltration/privilege escalation/instruction override/persistent manipulation) and attacker goal (confidentiality/integrity/availability). Seven variants would be new additions. Second, we introduce the PromptFort detection framework, which is a layer of semantic similarity in the API gateway that scores the input/output for LLMs against a constantly updated Prompt Injection Manifold and applies a tiered policy for blocking, flagging, and logging. Third, we present the challenge of PFB-1200, which includes 1200 attacks on all 12 variants, 4 LLM families and 3 API types. By addressing these challenges, PromptFort increases the macro-F1 of its performance to 93.5%, lowers the multi-hop attack success rate down to 6.4%, and can run at 7.2 ms p95 gateway latency, which is superior to all six baselines with production SLA requirements fulfilled.

Keywords: Adversarial attacks, LLM security, Large language models, Prompt injection

Cybersecurity today sits at an inflection point into becoming its own field. Developing endogenous pedagogical theory that can be externally identified as such and borrowed by other fields would yield a significant argument towards that becoming. Thirty years of The Colloquium for Information Systems Security Education (CISSE) has developed a rich corpus that we argue is necessary and sufficient, including deep conversations around first-order and second-order questions.

In this paper, we introduce Propagative Pedagogy, a first field-level endogenous vocabulary for cybersecurity education. Synthesizing concepts of Epistemic Cultures, Generative Justice, and multi-sitedness, Propagative Pedagogy recasts CISSE as a multi-sited epistemic culture in its own right, capable of producing knowledge and theory within and across sites that recirculates and amplifies value for all participants. The theory allows us to pose and answer a third-order question: "who produces cybersecurity pedagogy and what happens when the answer includes students themselves?" We answer by showcasing that authority and legitimacy are constantly reified by uptake—which recasts students as legitimate producers of pedagogy who already hold the authority to do so.

We provide a trio of vignettes from our research center to demonstrate the theory’s illustrative power: taking what appears to be three discrete projects and conceptualizing them as a meaningful multi-generational whole.

Propagative Pedagogy epistemically unifies legitimacy, authority, rigor, and value around individual research labs, cross-university collaborations, and the global CISSE network. It harmonizes the core values of cybersecurity into its pedagogy.

Keywords: cybersecurity, pedagogy, epistemic cultures, generative justice, multi-sitedness, science and technology studies

This study examines how entry-level cybersecurity work is represented in a Southwestern U.S. regional labor market and how employers characterize readiness for these positions. Situated within a research-practice partnership, the study draws on two data sources: 82 regional job postings collected across two time periods and four semi-structured interviews with technical leads involved in hiring, supervising, or training early-career cybersecurity professionals. Job postings analysis was grounded in the NICE Cybersecurity Workforce Framework and the U.S. Department of Education Employability Skills Framework, while interview data were analyzed qualitatively. Study findings suggest that K-12 cybersecurity pathways should prepare students for multiple career entry points by integrating broad technical foundations, employability skill development, and emerging AI literacy responsive to changing regional workforce needs.

Keywords: K-12 Education Pathways, Cybersecurity, Entry-level Jobs

The persistent gap between cybersecurity theory and applied practice continues to limit student readiness for real-world risk management roles. Cybersecurity clinics have demonstrated success in addressing this gap by engaging students in authentic, service-oriented risk assessments for under-resourced organizations. Traditional clinic models require sustained community partnerships, faculty capacity, and operational maturity that are difficult to scale. This paper presents an AI-enhanced Clinic-in-a-Box program that uses generative AI to produce synthetic organizational profiles capable of structured cybersecurity risk assessment learning. The program evolved from automated n8n workflow pipelines to direct Claude API integration, enabling richer and more contextually accurate profiles. These profiles encompass mission and governance, information technology environments, threat landscapes, and operational constraints, grounding student activities in sector-specific realism without the logistical and legal challenges of live client engagements. The program environment provides a guided workflow to support student learning through the applied risk management experiential learning process. The framework is aligned with the NICE Framework and supports deployable learning across tabletop exercises and hybrid cyber-range environments.

Keywords: experiential learning, cybersecurity clinics, AI-generated scenarios, risk assessment pedagogy, workforce development

Cloud Security Operations Centers (SOCs) can generate as many as 500,000 security alerts every day, but empirical research suggests that 70% of them are never investigated because analysts are overloaded. While automated triage systems have been constructed on rule-based Security Orchestration, Automation, and Response (SOAR) playbooks and traditional machine-learning classifiers, they each suffer from significant structural problems: they generate black-box, non-auditable results without maintaining a chain of reasoning that meets compliance requirements in regulated areas like PCI-DSS, HIPAA, SOC 2, and ISO 27001. More recently, the use of general-purpose Large Language Models (LLMs) for security triage has added a new failure mode: hallucination, as the Model produces factually inconsistent or fictional verdicts that can lead to overpriced and unnecessary incident responses or, even worse, to the underreporting of actual threats. We introduce SecOpsGPT, a multi-agent LLM framework to enable explainable and autonomous threat triaging in the cloud. SecOpsGPT consists of 3 cooperating specialized agents: (i) a Triage Agent that uses the chain-of-thought method of severity reasoning to classify incoming alerts, (ii) an Evidence Agent that retrieves and correlates supporting evidence using Retrieval-Augmented Generation (RAG) via MITRE ATT&CK, CISA KEV, and live OSINT sources, and (iii) a Verdict Agent that synthesizes a structured, compliance-ready triage verdict, along with a novel Reasoning Verifiability Score (RVS). A Hallucination Detection Layer validates all the LLM-generated verdict premises based on the retrieved evidence before any action is taken by the analysts.

SecOpsGPT demonstrates a macro-F1 of 91.3% ± 0.6 on a snapshot of 50,247 anonymized AWS CloudTrail and GuardDuty alerts labeled by human SOC analysts—a higher score than single-prompt GPT-4 (78.4%), IBM QRadar SOAR (71.2), or rule-based SIEM correlation (64.8). SecOpsGPT can also lower the Mean Time to Triage (MTTT) by 83% compared to the Mean Time to Triage of the most senior analyst, compared to the standalone GPT-4 baseline and generate machine-readable audit logs with 97.4% completeness compared to 14.7% for standalone GPT-4. The findings will form a foundation for next-generation autonomous SOC operations that is effective, verifiable, and compatible with compliance requirements for SecOpsGPT.

Keywords: Cloud threat triage, Multi-agent systems, large language models, Security operations center

The market for artificial intelligence (AI) penetration testing tools has expanded rapidly, yet vendor claims of autonomous operation are difficult for students and practitioners to evaluate, and structured classroom evidence about which tools deliver value remains scarce. We report on a two-assignment course activity in which 45 students in an upper-division offensive-security course each selected a distinct AI-assisted security tool, proposed a test plan, and then executed the tool hands-on in an isolated laboratory, producing 42 usable evaluations across 35 distinct tools. We designed the activity so that the class as a whole generated a cross-tool landscape dataset while individual students practiced critical evaluation of marketing claims, cost analysis, and comparison of AI-assisted testing against manual methods. The results showed that 16 evaluations (38 percent) reached success, 17 (40 percent) reached partial success, and 9 (21 percent) ended in failure, and that outcomes varied by tool category, with static code analysis appearing more mature and fully autonomous exploitation appearing more experimental. Cost, in the form of free-tier application programming-interface exhaustion, was the single most common limiting factor, affecting 11 of the 26 non-success evaluations. More than one student independently evaluated six of the tools, which let the class observe divergent outcomes for the same tool under different conditions. These findings suggest that a distributed, one-tool-per-student design is a low-cost way to give students hands-on evaluation experience while producing an empirically grounded snapshot of a fastmoving tool landscape.

Keywords: cybersecurity education, penetration testing, artificial intelligence, large language models, project-based learning, tool evaluation

K–12 schools increasingly rely on digital systems that support instruction, communication, assessment, and student services, making cybersecurity disruptions consequential for both organizational operations and classroom practice. Teachers occupy a distinctive position within school cybersecurity systems, yet little is known about how they interpret that role through professional experience, particularly after a disruptive cyberattack. This phenomenological case study explored the lived experiences of five K–12 teachers whose instruction was impeded by a district-wide cyberattack and examined whether and how those experiences influenced their subsequent perceptions of their role in school cybersecurity. Guided by sensemaking theory, analysis revealed five themes: navigating uncertainty amid limited information, dependence on digital systems, sustaining instruction through collective adaptation, recognizing shared vulnerability, and shifting perceptions of cybersecurity responsibility. The findings suggest that the cyberattack made previously hidden technological dependencies and vulnerabilities personally relevant while revealing teachers’ role in sustaining instructional continuity during disruption. Through reflection during the interviews, participants connected cybersecurity to responsibilities that extend beyond compliance, including student protection, responsible digital citizenship, instructional continuity, and care. These findings support a human-centered, sociotechnical view of K–12 cybersecurity in which teachers are not technical responders but protective actors who can contribute to organizational resilience when provided with relevant preparation, actionable guidance, and organizational support.

Keywords: K–12 cybersecurity, human factors, sensemaking, teachers, cyber resilience