Responsible Artificial Intelligence (AI) for Insider Threat September 2025 Office of International Nuclear Security This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor Battelle Memorial Institute, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof, or Battelle Memorial Institute. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. PACIFIC NORTHWEST NATIONAL LABORATORY operated by BATTELLE for the UNITED STATES DEPARTMENT OF ENERGY under Contract DE-AC05-76RL01830 Printed in the United States of America Available to DOE and DOE contractors from the Office of Scientific and Technical Information, P.O. Box 62, Oak Ridge, TN 37831-0062; ph: (865) 576-8401 fax: (865) 576-5728 email: reports@adonis.osti.gov Available to the public from the National Technical Information Service 5301 Shawnee Rd., Alexandria, VA 22312 ph: (800) 553-NTIS (6847) email: orders@ntis.gov Online ordering: http://www.ntis.gov RESPONSIBLE ARTIFICIAL INTELLIGENCE FOR INSIDER THREAT Pacific Northwest National Laboratory September 2025 Prepared by Jessica Baweja Jonathan Barr Chantell Murphy Pacific Northwest National Laboratory Richland, Washington 99354 CONTENTS Abbreviations, Acronyms, and Initialisms ..................................................................................................vii Executive Summary ................................................................................................................................... viii 1. Introduction ..................................................................................................................................... 1 1.1 Background ........................................................................................................................ 1 1.2 Current Study ..................................................................................................................... 3 2. Methodology .................................................................................................................................... 4 2.1 Literature Review and Capability Mapping....................................................................... 4 2.2 Taxonomy Development .................................................................................................... 4 2.3 AI Risk Framework Application .......................................................................................... 4 3. Results ............................................................................................................................................. 6 3.1 Identity and Document Verification Systems ................................................................... 6 3.1.1 Key Takeaways for Identity and Document Verification Systems................... 8 3.2 Trustworthiness Systems .................................................................................................. 9 3.2.1 Key Takeaways for Trustworthiness Systems ............................................... 12 3.3 Behavior Observation Systems ...................................................................................... 13 3.3.1 Key Takeaways for Behavior Observation Systems ...................................... 15 3.4 Impairment Detection Systems...................................................................................... 16 3.4.1 Key Takeaways for Impairment Detection Systems ..................................... 18 3.5 Access Control and Patrolling Systems ......................................................................... 19 3.5.1 Key Insights for Impairment Detection Systems .......................................... 21 3.6 Nuclear Material Accounting and Control Systems ...................................................... 21 3.6.1 Key Takeaways for Access Control and Patrolling Systems ......................... 23 3.7 Engagement Materials and Fact Sheets ....................................................................... 24 4. Conclusion .................................................................................................................................... 25 4.1 Implementation Considerations and Organizational Readiness ................................. 26 4.2 Caveats and Limitations ................................................................................................. 26 5. References ................................................................................................................................... 28 TABLES Table 1: Identity and Document Verification Systems Overview ............................................................... 6 Table 2: Implementation Checklist for Identity and Document Verification Systems ............................. 7 Table 3: Trustworthiness Systems Overview .............................................................................................. 9 Table 4: Responsible Implementation Checklist for Trustworthiness Systems .................................... 10 Table 5: Behavioral Observation Systems Overview............................................................................... 13 Table 6: Responsible Implementation Checklist for Behavior Observation Systems ........................... 14 Table 7: Impairment Detection Systems Overview ................................................................................. 16 Table 8: Responsible Implementation Checklist for Impairment Detection Systems .......................... 17 Table 9: Access Control and Patrolling Systems Overview ..................................................................... 19 Table 10: Responsible Implementation Checklist for Access Control and Patrolling systems ............ 20 Table 11: Nuclear Material Accounting and Control Systems Overview ............................................... 21 Table 12: Responsible Implementation Checklist for NMAC Systems .................................................. 22 ABBREVIATIONS, ACRONYMS, AND INITIALISMS AI Artificial Intelligence DIVA Deep Intermodal Video Analytics HR Human Resources IAEA International Atomic Energy Agency IARPA Intelligence Advanced Research Projects Activity IT Information Technology ITM Insider Threat Mitigation LLM Large Language Models ML Machine Learning NIST National Institute of Standards and Technology NMAC Nuclear Material Accounting and Control PNNL Pacific Northwest National Laboratory SALT Scalable and Automated Linking Technology SVM Support Vector Machines TSA Transportation Security Administration UAS Unmanned Aerial Systems EXECUTIVE SUMMARY This report examines the application of artificial intelligence (AI) technologies for insider threat mitigation (ITM) programs in nuclear security facilities. Insider threat detection presents unique challenges due to the subtle and adaptive nature of these threats, the complex signatures involved, and the scarcity of available data for analysis. Traditional human-centered approaches, while essential, face limitations in processing large amounts of data continuously and detecting subtle patterns across multiple systems. AI technologies can potentially address these limitations by providing 24/7 monitoring capabilities, identifying complex patterns that might escape human observation, and offering consistent application of security criteria. However, the deployment of AI in nuclear security contexts introduces significant new risks, including workflow disruption, expanded attack surfaces, potential for misuse, and ethical concerns regarding privacy, fairness, transparency, safety, and security. The high-consequence nature of nuclear security decisions demands careful consideration of these risks and systematic approaches to their mitigation. This analysis applies the US National Institute of Standards and Technology (NIST) AI Risk Management Framework to evaluate AI applications for ITM. The framework's core principles— accountability, fairness and bias mitigation, privacy and data protection, security and abuse prevention, explainability and interpretability, safety, and performance—provide an inclusive lens for assessing potential risks and developing mitigation strategies. The research involved four primary phases: literature review and capability mapping, taxonomy development, AI risk framework application, and implementation guidance development. The analysis identifies six functional categories of AI applications for nuclear security ITM: 1. Identity and Document Verification Systems: These systems authenticate individuals and validate documents by comparing credentials, biometric features, or documents against trusted records. Applications include facial recognition for identity verification, document-based identity verification, and fraud detection for document verification. These technologies serve as verification tools during background checks, visitor processing, or personnel onboarding while maintaining human-in-the-loop decision-making. 2. Trustworthiness Systems: These systems assess personnel risk levels by evaluating patterns across historical records and identifying potential concerns. They aggregate information from multiple sources to generate comprehensive risk assessments and flag concerning patterns requiring human investigation. Applications include automated risk scoring from documents and automated issue identification in background records. 3. Behavior Observation Systems: These systems detect anomalous activities that might indicate insider threats by establishing baselines of normal behavior and flagging significant deviations. They analyze patterns in physical movement, facility access, network activity, and digital communications. Applications include fitness-for-duty monitoring, video analytics systems, and cyber behavior analysis. 4. Impairment Detection Systems: These anomaly detection systems identify potential fitness-for- duty concerns by analyzing physiological and behavioral indicators of fatigue, intoxication, or other impairment. Applications include fatigue detection through multi-modal inputs and drug/alcohol impairment detection through behavioral pattern analysis. 5. Access Control and Patrolling Systems: These systems secure facility boundaries by authenticating personnel, screening for prohibited items, and conducting autonomous security monitoring—useful in multiple security applications, including detection of insider threats. They make real-time decisions that directly impact facility security, often with limited human intervention in the immediate decision loop. Applications include biometric access controls, automated screening at security checkpoints, and autonomous security patrols. 6. Nuclear Material Accounting and Control Systems: These systems detect deviations in nuclear materials accounting data that may indicate theft or diversion. Applications include material movement characterization, transaction anomaly detection, and synthetic data generation for training purposes. Organizations must carefully assess their readiness across multiple dimensions before implementing AI for ITM. Technical readiness requires appropriate data quality, information technology (IT) infrastructure, and organizational capacity to integrate AI systems without creating vulnerabilities. Organizational readiness encompasses policy frameworks, personnel training, legal compliance, and coordination across multiple organizational functions. All AI applications for ITM should be implemented as decision support tools that enhance human capabilities rather than replace human judgment. The high-consequence nature of nuclear security decisions, combined with current AI limitations and sparse insider threat data, makes human oversight essential for responsible implementation. Organizations should develop roadmaps that deploy AI applications based on organizational readiness and security benefits rather than implementing systems in isolation. Identity and document verification systems, for example, build upon established commercial technologies with clear implementation pathways. Behavior observation systems require more sophisticated infrastructure and extensive policy development. This framework provides a foundation for risk assessment but cannot address all potential concerns for every organizational context. The rapid pace of AI advancement means this analysis represents a snapshot of current capabilities that will require regular updates. Organizations must conduct their own thorough assessments of how AI applications interact with their specific operational procedures, security measures, and regulatory requirements. The commercial AI landscape is highly dynamic, and the specific examples referenced should be understood as illustrative rather than endorsements. Organizations need specialized technical, legal, and policy expertise throughout implementation processes and should work closely with regulatory bodies to ensure compliance. AI applications offer significant potential to enhance nuclear security ITM programs through capabilities that extend beyond traditional approaches. However, successful implementation requires careful planning, systematic risk management, and ongoing commitment to monitoring and adaptation. Organizations that invest in developing strong foundational capabilities for AI risk management will be better positioned to realize these benefits while maintaining the highest standards of nuclear security. The framework presented here provides a structured approach to evaluating and implementing AI technologies for ITM while maintaining appropriate human oversight and addressing inherent ethical and operational challenges. As AI technologies continue to evolve, the procedures and frameworks established today must be regularly updated to reflect changing capabilities, threats, and regulatory requirements. Most critically, responsible AI implementation for ITM is not a one-time effort but an ongoing process requiring continuous evaluation, adaptation, and unwavering commitment to safety and security principles in the nuclear domain. CHAPTER 1 1. INTRODUCTION Insider threat detection and mitigation remains a pressing challenge in nuclear security, given the complex signatures involved, the paucity of data available, and the relatively subtle and adaptive nature of insider threats. Artificial intelligence (AI) enabled technologies provide a promising new method for combatting detection of insider threats, potentially allowing for more rapid intervention and protective action. Recent work has explored the use of AI for insider threat in a variety of different ways, including through anomaly detection, natural language processing supported by large language models (LLMs), graph-based approaches, and user behavior analytics [1-5]. The introduction of AI technologies into the nuclear security enterprise, however, also introduces additional risks in the form of workflow disruption, broadening of the attack surface, and increased potential for misuse [6-11]. Furthermore, the application of AI in the high-consequence domain of nuclear security either creates or exacerbates ethical concerns regarding privacy, fairness, transparency, safety, and security of AI enabled systems [12-14]. These concerns have led to a variety of efforts by the United States Government to introduce best practices and guidance for the use of AI to help ensure that it is deployed in a secure, effective, and responsible manner [9, 15]. In this report, we present a review of the current and potential future uses of AI technology for insider threat mitigation (ITM) programs in nuclear security and develop a framework to address the risks to the organization when applying AI to ITM. We apply a broad definition of AI, including any system that can generate outputs such as predictions, recommendations, or decisions at various levels of autonomy [9]. This includes expert- or rule-based systems, machine learning (ML) systems, and more recent AI systems based on deep learning technologies, such as LLMs. We focus on AI systems that could be used in support of ITM through preventive, protective, or comprehensive measures as described by the International Atomic Energy Agency (IAEA), rather than through more general administrative support (e.g., automation of incident reports using generative AI) [16]. Using the information gathered in this review, we leverage existing knowledge on safety, security, and ethical risks of AI technologies to evaluate the risks presented by applying AI to nuclear security ITM programs. Finally, we discuss mitigation approaches for managing the risks associated with AI deployment for ITM programs. 1.1 BACKGROUND Prevention and detection of insider threats is a complex and multidisciplinary problem, and research into more effective mitigation measures has now spanned decades [17]. Despite dedicated research, insider threat remains a critical challenge when seeking to protect nuclear facilities and material from the risk of sabotage or other threats. It remains true that, “virtually all the cases of nuclear theft in which the circumstances are known were perpetrated either by insiders or with the help of insiders,” underscoring the need for effective approaches to managing the risk of insider threat [18]. Much of the challenge results from the fact that insiders have authorized access and, in many cases, detailed knowledge of facilities and security measures that increases the likelihood of attack success [16-18]. This can make detection, delay, and response to those attacks more challenging. Indications of potential insider attacks may be subtle, spread across multiple systems, and difficult for humans (which may include physical security specialists, cybersecurity professionals, human resources personnel, and supervisors) to identify and appropriately respond [19-21]. 1 It is no surprise, then, that researchers have explored the use of “big data,” ML, and AI approaches to help predict or detect insider threats [1-5, 22]. These technologies have key advantages when addressing the challenge of insider threat: they can process vast amounts of data, detect subtle and complex patterns—even across sensors—and they can monitor systems and facilities 24/7 without fatigue. Along with these advantages, however, come significant challenges that need to be addressed. First, it is critical that systems applied in high-consequence domains, such as the protection of nuclear material and facilities, perform effectively. AI system failure in nuclear security could have devastating impacts on employee and public safety and security, making it critical to ensure that the systems perform well at their intended functions. However, evaluation of AI system performance is an active area of research and remains challenging. This is partially due to a need for clear regulation and standards regarding the required performance of a system for high-consequence use [12, 23, 24]. System performance evaluation is especially difficult in the domain of insider threat, where insider threat training data are extremely sparse due to the small number of known incidents. As a result, practical evaluation of AI systems for insider threat detection remains extremely challenging, forcing the use of proxy metrics that may not generalize to real-world performance [1]. Recent collaboration between the Canadian Nuclear Safety Commission, the UK Office for Nuclear Regulation, and the US Nuclear Regulatory Commission has explored many of these considerations, and work is emerging to outline the regulations and standards needed to deploy AI systems in nuclear applications [23, 24]. Researchers have also explored “affirmative safety” approaches, which would ask developers to take proactive steps to demonstrate the performance and safety of their systems and outlining organizational practices to support safety [12]. While the standards and approaches are developing, organizational methods to assess the performance and safety of AI systems in a systematic, proactive, and holistic way—and identify failure modes and backup systems for those systems—is a critical component of deploying them in a responsible and effective manner. Second, maintaining engaged human oversight of the AI system, and accountability for its outcomes, is key to ensuring that humans remain responsible for the security of the facility. There is a large body of evidence that humans may over-rely on AI systems, trusting or using system outputs even when they may not be accurate [25-28]. Relatedly, many AI systems—especially deep learning systems—lack transparency, meaning that the recommendations or decisions made by the system may not be readily explainable to a human user [11, 29-33]. The lack of sufficient transparency or explainability makes it difficult for users to understand when a system should—or should not—be relied upon. As a result, when deploying AI systems for nuclear security, it must be clear to end users how the system results should be used, how they should be verified, and when they cannot or should not be relied upon. Because the consequences of AI systems being applied to insider threat are very high (either security lapses in the case of false negatives or reputational and employment harms for false positives), human understanding of AI system outputs is paramount. Operators must, at every stage, be capable of explaining why any decisions that affect human safety, security, or livelihoods were made. AI systems also introduce risks to privacy and security that need to be considered. The aggregation of data—especially personally identifying information on applicants or employees—for AI system training or analysis may introduce or exacerbate concerns regarding privacy [34, 35]. In addition, AI systems may inadvertently release private information through a phenomenon known as training data leakage, whereby entire segments of training data are regurgitated in system outputs [36]. Depending on the country where the system is deployed, there are complex regulations that may need to be navigated regarding the protection of private information that need to be considered when deploying AI systems for nuclear security. AI systems also expand the attack surface, introducing new threats that need to be considered, such as the threat of adversarial AI [37, 38]. 2 Finally, AI systems may have problems with bias—both in terms of demographic characteristics when applied to people, and, in more technical terms, whereby the model demonstrates variable performance across different classes, features, or in different contexts [39, 40]. Again, the sparsity of data and signatures for insider threats may make this problem especially concerning for AI systems for ITM. Without representative, relevant, accurate, and generalizable data, issues of data or demographic bias may emerge when these systems are deployed, resulting in inaccurate, unfair, or inconsistent performance. Organizations must, therefore, consider the quality of the data on which the models were trained and evaluate the performance of the system under variable conditions to ensure that it meets reasonable standards for fair, unbiased, and robust performance. Critically, regardless of where they are applied, AI systems should not be used for prediction of insider threat. Given the rarity of the event, the sparsity of the data, and the severity of false positives, prediction is not only unlikely to be accurate, it is also highly likely to produce damaging errors. AI systems can and should serve only as providers of information for professional evaluators (e.g., adjudicators) to make determinations regarding trustworthiness. Especially given the “black box” nature of many AI approaches, this is a critical boundary for responsible AI implementation in insider threat. 1.2 CURRENT STUDY All of these issues emphasize the need for a framework to help organizations manage the risks of AI systems as they seek to realize their capabilities for ITM. In this study, we seek to address this need by conducting a review of the ways that AI has been or could be applied to insider threat and creating a taxonomy of those applications. Then, we apply principles from the US National Institute of Standards and Technology (NIST) AI Risk Management Framework to understand the risks of that application of AI and to identify organizational processes to mitigate them [9]. We apply this framework because a review of existing ethical frameworks and approaches suggests that it represents an emerging consensus regarding the ethical principles most relevant to effective, reliable, and trustworthy use of AI technologies: that they perform well at their tasks (i.e., valid and reliable), and are safe, secure, privacy-enhanced, fair, accountable, and explainable. 3 CHAPTER 2 2. METHODOLOGY The methods applied in this work consisted of four primary phases: literature review and capability mapping, taxonomy development, risk assessment framework application, and implementation guidance development. Each phase is described in greater detail in the sections below. 2.1 LITERATURE REVIEW AND CAPABILITY MAPPING We conducted a review of academic literature, gray literature, and commercial AI capabilities relevant to ITM. This review included peer-reviewed research papers, government reports, industry white papers, and commercially available AI-powered security products. The search focused on applications of AI, ML, and automated systems for detecting, preventing, or mitigating insider threats across various high-security domains. To provide structure to this analysis, we mapped identified AI capabilities to the International Atomic Energy Agency's framework for ITM programs as outlined in Preventive and Protective Measures Against Insider Threats (Nuclear Security Series 8-G) [16]. This mapping process classified each AI application according to whether it supported preventive measures (designed to reduce the likelihood of insider threats), protective measures (designed to detect and respond to insider threats), or comprehensive measures (integrated approaches combining multiple elements). 2.2 TAXONOMY DEVELOPMENT Using the information gathered through the literature review, we developed a functional taxonomy to categorize AI applications for ITM. The taxonomy development process involved iterative exploration of multiple organizational approaches, with the research team evaluating different categorization schemes based on their ability to capture the essential characteristics of AI technologies, their operational contexts, and their associated risk profiles. Through collaborative analysis and discussion among team members, we converged on a taxonomy organized around six functional categories based on the primary security objectives and operational characteristics of different AI applications: (1) Identity and document verification systems, (2) Trustworthiness systems, (3) Behavior observation systems, (4) Impairment detection systems, (5) Access control and patrolling systems, and (6) Nuclear material accounting and control systems. Notably, these applications are somewhat interrelated. Trustworthiness assessments are the end goal of most personnel security focused approaches; generally, fitness-for-duty evaluations are conducted in service of a trustworthiness assessment. Behavior observation systems, generally, provide data as inputs to either fitness-for-duty evaluations or for trustworthiness assessments more broadly. However, there are AI models and systems that have targeted each level of this hierarchy (from broadly evaluating trustworthiness to granular behavior monitoring); thus, we present these different applications separately. 2.3 AI RISK FRAMEWORK APPLICATION We applied principles from the NIST AI Risk Management Framework to systematically assess the risks associated with each category of AI applications [9]. The NIST framework's core principles of trustworthy AI systems—Accountability, Fairness and Bias, Privacy and Data Protection, Security and 4 Abuse Prevention, Explainability and Interpretability, Safety, and Performance (called Validity and Reliability in their approach)—provided a comprehensive lens for evaluating potential risks and challenges. For each functional category, team members conducted independent risk assessments, examining how each NIST principle applied to specific characteristics and use cases within that category. These individual assessments were then compared and discussed among the research team to identify areas of agreement and resolve any differences in risk perception through consensus-building discussions. For each functional category of AI applications, we developed implementation guidance made up of three components. First, we created descriptions of specific technology applications within each category, explaining their core functions, typical algorithms employed, and the fundamental security questions they address. Second, we identified prerequisites for implementation, including necessary infrastructure, policies, and organizational capabilities required for successful deployment. Third, we created "Responsible Implementation Checklists" that translate each NIST AI Risk Management Framework principle into specific organizational processes and controls tailored to the characteristics of each AI application category. To support practical application of this research, we developed supplementary materials designed for INS partners. These included engagement materials formatted for presentation to international audiences and fact sheets summarizing key information for each group of AI technologies. These materials distill the technical analysis into accessible formats suitable for policymakers, security professionals, and organizational decision-makers considering AI adoption for ITM. 5 CHAPTER 3 3. RESULTS The results of this analysis showcase a large landscape of AI applications for ITM, each offering distinct capabilities with unique implementation challenges and risks. The results are organized according to our six-category functional taxonomy, with each category addressing specific security objectives and operational requirements within nuclear facilities' ITM programs. Throughout, we provide a description of the technologies in each domain and implementation guidance to address risks identified. We provide academic or commercial examples of relevant technologies in each section; note that this is not an endorsement of those technologies. This information is provided only to illustrate that the categories and descriptions apply to products currently available on the commercial market or to capabilities actively being researched. 3.1 IDENTITY AND DOCUMENT VERIFICATION SYSTEMS Identity and document verification systems employ AI to authenticate individuals and validate Key Benefits documents by comparing presented credentials, ✓ Faster document biometric features, or documents against trusted processing records. These systems analyze identity documents ✓ Consistent evaluation for consistency and authenticity, match facial criteria features to verified images, and detect potentially ✓ Improved detection of fraudulent documentation. In nuclear facilities, concerns these technologies may serve as verification tools ✓ Scalable for high-volume during background checks, visitor processing, or processing personnel onboarding. They can help determine whether someone is who they claim to be and whether their documentation is legitimate while maintaining a human-in-the-loop approach for final verification decisions. Table 1 shows different applications of AI in this domain as well as the central question that each system addresses, how it is generally applied, common algorithm types, and the nature of the mathematical problem being addressed (e.g., classification, clustering, regression). We provide brief descriptions of each of the technologies in the sections below. Table 1: Identity and Document Verification Systems Overview Technology Central Question Application Common Algorithm Mathematical Application Type (What it's Types (How it's Built) Problem Type trying to do) (How it Works) Document-Based Does this document Document Random Forest Classification Identity belong to the applicant Processing Classifier, Support Verification under investigation? Vector Machines (SVM), Neural Networks Fraud Detection Is the document under Document Neural Networks, Classification for Document review fraudulent? Processing Random Forest Verification Classifier Facial Does this image Computer Neural Networks Classification Recognition for belong to the applicant Vision under investigation? 6 Identity Verification AI-powered facial recognition technology supports background check processes by comparing applicant photographs against images in existing records (such as driver's licenses or criminal records) to verify identity. The system generates confidence scores or classifications indicating the likelihood of a match between the provided photograph and images in official records, serving as one component of identity verification during background checks. One of the most well-known and widely-used AI-powered facial recognition software is the Clearview AI app, which has been used by law enforcement across the US to locate suspects in criminal investigations [41]. AI-powered identity verification systems support background check processes by automating the matching of records across databases to identify information relevant to a specific individual. The system generates confidence scores or classifications to indicate whether different records likely belong to the same person, serving as a preliminary screening tool that helps human analysts more efficiently review and verify background information for employment decisions. There are several examples of this automated identity verification in commercial products, such as LexisNexis’ Scalable and Automated Linking Technology (SALT), Veriff Identity Verification, or Checkr’s AI- powered background checks [42-44]. AI-powered fraud detection supports background check processes by automatically analyzing submitted documents for potential manipulation, forgery, or inconsistencies. The system examines various records (employment, financial, educational, etc.) and flags potential authenticity concerns for human review, serving as an initial screening tool to enhance document verification processes. AI use in fraud detection has increased in recent years in industry and its use has been explored by the US Treasury Department to understand the regulatory gaps that need to be addressed for effective implementation [45-47]. In order to implement AI-enabled identity and document verification systems, organizations must meet certain prerequisites. First, they must have a digital document management system with security measures in place to appropriately protect the sensitive information they contain. Relatedly, they must have or create clear document handling and system access procedures to protect applicant privacy and to prevent misuse of the system. Finally, before implementing an AI system for identity or document verification in background check processes, the organization must consider how this system will integrated with their current human resources (HR) and security process—that is, how will information be shared with appropriate parties and who will act upon information identified? Table 2 presents the risk considerations for AI systems supporting identity and document verification as well as organizational processes that can be used to manage that risk. For these systems specifically, many of these risks are related to the sensitivity of the information collected (and protecting that information appropriately) and verifying the results prior to use in the applicant decision-making process. It is also critical to evaluate the performance of these AI systems both prior to procurement and during deployment to ensure that they continue to perform at sufficient levels across demographic and other groups. Table 2: Implementation Checklist for Identity and Document Verification Systems Principle Key Considerations Managed Process through… Accountability Personnel must remain Human Implement verification processes accountable for system Oversight where system results are reviewed by 7 Principle Key Considerations Managed Process through… decisions, with the system trained personnel before making serving as a supporting tool determinations about identity or rather than an autonomous document authenticity. decision-maker. Fairness & These systems may exhibit Alternative Maintain and regularly use multiple Bias varying performance across Verification verification approaches to demographic groups or Methods complement AI systems, ensuring document types. individuals aren't disadvantaged by technological limitations. Privacy & These systems process Secure Data Establish comprehensive protocols Data sensitive personal information Handling for the secure collection, storage, Protection including facial images and transmission, access, and disposal of identity documents. identity documents and biometric data. Security & These systems may be Access Controls Restrict system access to authorized Abuse misused through unauthorized personnel with legitimate business Prevention searches, inappropriate needs, implementing role-based access, or attempts to permissions and comprehensive manipulate verification audit logs. outcomes. Explainability The system must provide Documentation Maintain records of all verification & indications for why the results Standards decisions, including system results, Interpretability are flagged, and personnel human determinations, and must be able to understand rationales for decisions that deviate the role and limitations of the from system recommendations. system. Safety These systems must support Clear Establish documented protocols for rather than impede effective Procedures handling system limitations, background checks and uncertain results, outages, and procedures should avoid escalation paths for concerning personnel overreliance. findings. Performance These systems should Performance Prior to procurement and during demonstrate reliable Monitoring deployment, assess system accuracy, performance in flagging false positive/negative rates, and documents or images for performance across different review. demographic groups and document types. 3.1.1 KEY TAKEAWAYS FOR IDENTITY AND DOCUMENT VERIFICATION SYSTEMS • These systems are tools, not final decision makers, • Regular training and audits maintain system effectiveness and fairness, and • Maintaining human in the loop is essential for responsible use. 8 3.2 TRUSTWORTHINESS SYSTEMS Trustworthiness systems use AI to assess personnel risk levels by evaluating patterns across historical records and identifying Key Benefits potential concerns. Here, trustworthiness refers ✓ Consistent application of to the individual’s ability to safeguard sensitive evaluation criteria material and information. These systems ✓ More efficient aggregate information from multiple sources to identification of potential generate comprehensive risk assessments, flag concerns concerning patterns, and identify specific issues ✓ Enhanced pattern requiring human investigation. In nuclear recognition across large facilities, these technologies enhance personnel volumes of information security programs by providing more consistent evaluation of complex information sets when making trustworthiness determinations. They serve as decision support tools that help security professionals identify patterns of concern that might be missed through traditional evaluation methods. Table 3 provides an overview of the ways that AI systems have been applied to the problem of evaluating personnel for trustworthiness—which, in the existing products, is generally described as personnel risk. This includes an overall risk score for applicants or employees as well as automated identification of issues (i.e., concerns) found in the individual’s background. Table 3: Trustworthiness Systems Overview Technology Central Question Application Type Common Mathematical Application (What it's trying to Algorithm Types Problem Type do) (How it's Built) (How it Works) Risk Scoring What is the overall Predictive/Scoring Linear Regression from trustworthiness level of Regression, Documents this individual based on Random Forest, their records? Neural Networks Automated Are there specific Document Neural Networks, Classification Issue concerns in this person's Processing SVM Identification background that warrant further investigation? AI-powered risk scoring systems support background check processes by automating the document review and evaluation process to assign a risk score that reflects a person’s likelihood to be of concern. The system aggregates and reviews criminal records, financial statements and reports, employment history, and other records to generate an overall risk score indicating the trustworthiness of the applicant. There are several examples of automated risk scoring in commercial application. Airbnb uses a proprietary algorithm to evaluate customers' risk for property damage or risk [48]. In recent years, researchers have explored the use of machine learning methods for credit scoring [49]. Financial technology companies have also begun to use “alternative data” (beyond those traditionally used in credit scoring). For example, some lenders use this so-called alternative data for credit decisions and ML approaches to help make more accurate lending decisions [50]. However, the use of machine learning—and especially deep learning—has raised concerns regarding transparency for decisions amongst consumers [51]. Nonetheless, there is evidence from the financial technology industry that risk scoring from documents may provide value when making consequential decisions about an applicant, suggesting promise for a risk scoring approach for ITM. Of course, the domain of financial 9 reliability is a much more limited one than overall risk or trustworthiness; additional research is certainly warranted to expand the application of AI systems from a single domain to a holistic judgment of the overall trustworthiness or riskiness of an individual. AI-powered background check systems automate the process of identifying potential concerns in applicant records by analyzing different data sources (criminal, financial, employment) and flagging relevant issues for human review. The system helps streamline the background check process by automatically identifying and categorizing potential concerns that warrant further investigation. Automated identification of concerns in background checks is used by many companies who also leverage AI for their identity resolution or verification processes, including Checkr [44]. In addition, automated detection of issues during employment is a cornerstone of the US government’s reformed clearance process, Trusted Workforce 2.0 [52]. In order to implement AI-enabled trustworthiness systems, organizations must meet several critical prerequisites. First, they must establish secure data integration capabilities that can safely aggregate information from multiple sources while maintaining appropriate access controls and data integrity. Second, organizations must define clear risk factors and evaluation criteria to ensure consistent and fair adjudication standards across all personnel assessments. Given that these systems process highly sensitive personal information, organizations must also implement robust data protection infrastructure that complies with applicable privacy regulations and security requirements. Finally, before deploying trustworthiness systems, organizations must develop a comprehensive plan for integrating these capabilities with existing personnel security processes— that is, how will automated risk assessments be incorporated into human decision-making workflows, who will have access to system outputs, and how will personnel act upon flagged concerns? The risk considerations for AI systems supporting trustworthiness assessments are particularly complex given the high-consequence nature of personnel security decisions and the potential for algorithmic bias in risk scoring systems. This is especially concerning given that the judgment of overall trustworthiness is fundamentally subjective, allowing for the potentially strong influence of bias. Furthermore, the lack of clear and objective data regarding insider threats makes the scientific evaluation of trustworthiness especially challenging. Many of the risks of AI-based trustworthiness systems therefore relate to ensuring fair and transparent evaluation processes while protecting the sensitive personal information these systems analyze. It is also critical to validate the performance of these AI systems both prior to procurement and continuously during deployment to ensure they maintain acceptable accuracy levels and do not exhibit bias across different demographic groups or risk categories. Note that the evaluation of performance is much more straightforward for automated issue identification (where there is a clearly verifiable right and wrong) than for automated risk or trustworthiness scoring. Thus, organizations should be especially critical of systems designed to measure personnel risk to ensure that the resulting scores accurately reflect the characteristics that they wish to evaluate in a verifiable manner and fairly across demographic groups. Information on how to address some of these risks at the organizational level is displayed in Table 4. Table 4: Responsible Implementation Checklist for Trustworthiness Systems Principle Key Considerations Managed Process through… Accountability These systems inform Decision Establish a formal framework consequential decisions about Authority that clearly defines who has personnel trustworthiness, Policy authority to make decisions 10 requiring clear oversight and based on system outputs, responsibility structures. with documentation requirements increasing with risk score severity. Fairness & Risk scoring and issue Adjudication Develop standardized Bias identification may reflect Standards procedures for evaluating historical biases in records or flagged issues and risk uneven availability of scores in proper context, with information across different explicit consideration of groups. potential information gaps or systemic factors. Privacy & Data These systems process highly Data Implement strict controls on Protection sensitive personal information Governance data access, usage, from multiple sources, creating retention, and sharing, with significant privacy concerns. formal protections against function creep or unauthorized expansions of system use. Security & Access to these systems could Access Create multi-level access Abuse enable inappropriate Controls restrictions based on need- Prevention investigations or misuse of to-know principles, with sensitive personal information. enhanced monitoring and audit requirements for high- risk functions, and implement comprehensive training for personnel on appropriate system use. Explainability & The system must provide Adjudication Maintain clear Interpretability sufficient information so that Transparency documentation of which personnel can understand what factors contribute to risk factors contribute to risk scores scores and issue or issue flags to make informed identification, ensuring decisions. decision-makers can explain the basis for system outputs. Safety Over-reliance on automated Multiple Implement complementary assessments could lead to Analysis assessment approaches that missed issues or inappropriate Methods don't rely solely on conclusions about automated analysis, with trustworthiness. manual review requirements scaling with risk level. Performance These systems must reliably Regular Conduct periodic testing with identify genuine concerns while Validation known outcomes to verify maintaining manageable false Testing system performance, with positive rates, which is notably more frequent and rigorous challenging in this application 11 given limited data. Performance validation required for high- is likely to be a significant (and risk applications. potentially insurmountable) obstacle. 3.2.1 KEY TAKEAWAYS FOR TRUSTWORTHINESS SYSTEMS • These systems should inform but never replace human judgement in personnel security decisions, • Clear documentation of factors contributing to risk assessment is essential for fair evaluation, and • Strong privacy controls are critical due to the sensitive personal information processed. 12 3.3 BEHAVIOR OBSERVATION SYSTEMS Behavior observation systems apply AI to detect anomalous activities that might indicate insider Key Benefits threats by establishing baselines of normal ✓ Continuous monitoring behavior and flagging meaningful deviations. beyond human capability These systems analyze patterns in physical ✓ Consistent application of movement, facility access, network activity, and detection criteria digital communications. In nuclear facilities, ✓ Integration of physical and these technologies provide continuous cyber indicators monitoring capabilities across physical and ✓ Early identification of cyber domains, helping security teams identify potential insider threats concerning behavior patterns that would be difficult to detect through manual observation alone. They operate as persistent monitoring tools that extend security teams' awareness while requiring human analysis of alerts before actions are taken. Table 5 presents an overview of three ways that AI has been applied to behavior observation systems. Table 5: Behavioral Observation Systems Overview Technology Central Question Application Type Common Mathematical Application (What it's trying to Algorithm Types Problem Type (How it do) (How it's Built) Works) Fitness for Duty Is this individual Anomaly Neural Networks, Clustering and Behavior displaying Detection K-means Observation concerning Clustering behavioral patterns that warrant further investigation? Location and Is this movement or Computer Vision Neural Networks Classification Anomalous activity pattern Behavior Tracking unusual or concerning for this individual or location? Cyber Behavior Does this digital Anomaly Neural Networks, Classification & Analysis activity indicate Detection SVM Clustering potential insider threats or system compromise? AI systems that support fitness for duty analysis and behavior observation monitor behavioral patterns and apply anomaly detection approaches to identify deviations from the norm, signaling potential security threats. In this sort of an application, the AI system analyzes surveillance footage, access control records, or other sensors and outputs alerts when anomalous behavior is detected on or across systems. There is overlap between this application of AI and systems that perform location or anomalous behavior in general. Here, the focus is primarily on identifying patterns and deviations from those patterns, whether or not those occur in a physical realm. Instead, they may involve only cyber domains or correlation between cyber and physical realms. Many of these approaches focus on cyber behavior, given the relative immediacy of those data for analysis. For example, the company 13 Darktrace uses behavioral analytics and AI to establish baselines and identify departures from typical patterns, such as unusual logins or data transfers [53]. In one recent example of anomaly detection in nonproliferation analysis, researchers used a machine learning algorithm on a digital twin of a research reactor to detect proliferation-related anomalies, achieving a classification accuracy of 99% [54]. In general, there is evidence in both the commercial market and scientific literature that AI-based behavior observation may provide some promise for detecting concerning anomalies [1, 2, 4, 55]. However, notably, the systems can detect only observable behaviors; manual observations based on human judgment, knowledge of the individual, or experience may provide additional information beyond the sensors that AI system cannot detect. Furthermore, even high classification rates can fail when applied to low base rate systems such as insider threat. Nonetheless, there is some indication that AI systems can provide value in conducting fitness for duty or behavior observations. AI-powered video analytics systems analyze surveillance feeds in real-time to detect potential security threats through identification of anomalous behavior, unattended objects, and other security-relevant events. In this use of AI, the system generates alerts for security personnel when it detects patterns that deviate from normal facility operations. These systems go beyond computer vision systems for intrusion detection or alarm, but focus on detection of typical operational patterns and deviations from them. There are a number of commercially available products that purport to have this capability, including Actuate, Verkada, and BriefCam [56-58]. Similarly, a previous Intelligence Advanced Research Projects Activity (IARPA) project called Deep Intermodal Video Analytics (DIVA) sought to develop technologies to detect specific activities in video feeds for forensic analysis or real-time alerts [59]. AI-powered monitoring and anomaly detection systems seek to identify and flag unauthorized transfers and prevent data loss. The system analyzes network traffic and digital communication and alerts cyber security professionals of potential data leakage, suspicious activity, or unauthorized transfers. The company Darktrace offers an AI-based anomaly detection system [53]. A variety of other commercial entities offer AI-based behavior analytics for security that could be applied in a nuclear setting [60]. Implementing AI-enabled behavior observation systems requires organizations to establish several foundational capabilities before deployment. Organizations must first develop clear definitions of baseline “normal” behavior patterns for their specific operational environment, as these systems rely on identifying deviations from established norms. The systems also require integration with existing physical and cyber monitoring infrastructure to avoid creating security gaps or redundant capabilities. The use of AI systems for behavior observation programs also introduces novel risks to privacy, accountability, and security that must be considered before and during implementation. Given the continuous monitoring nature of these systems, organizations must establish robust notification and escalation workflows that define how alerts will be handled, who will respond to different types of behavioral flags, and what actions may be taken based on system outputs. Finally, the privacy implications of continuous behavioral monitoring necessitates privacy impact assessments and appropriate legal review to ensure compliance with applicable regulations and organizational policies before system activation. Organizational processes for addressing these concerns are described in Table 6. Table 6: Responsible Implementation Checklist for Behavior Observation Systems 14 Principle Key Considerations Managed Process through… Accountability Behavior monitoring systems Alert Investigation Establish formal procedures for must operate within clear Framework reviewing and investigating alerts, authority structures with with clear documentation defined oversight requirements and escalation paths responsibilities. for different alert types. Fairness & Bias Monitoring systems may Diverse Human Implement multi-person review incorrectly flag cultural Review Teams protocols with diverse differences or neurodivergent representation to evaluate alerts behaviors as suspicious. within appropriate cultural and individual contexts. Privacy & Data Continuous monitoring Minimization & Define and enforce strict data Protection creates significant privacy Purpose collection limitations, focusing only concerns regarding personal Limitation on security-relevant behaviors and and potentially sensitive implementing technical controls information. against function creep. Security & Behavior monitoring systems Anti-Harassment Create technical and procedural Abuse could enable targeted Safeguards protections against targeting Prevention harassment, unauthorized specific individuals, with surveillance, or retribution. independent oversight of monitoring patterns and usage. Explainability & Security personnel must Alert Context Require systems to provide specific Interpretability understand what specific Documentation behavioral indicators that triggered behaviors triggered alerts to alerts, maintaining clear distinction make informed judgments. between observed behaviors and algorithmic inferences. Safety Over-reliance on automated Balanced Maintain traditional monitoring detection could create Monitoring methods alongside automated security vulnerabilities Approach systems, with regular exercises to through missed behaviors or ensure human observation alert fatigue. capabilities remain effective. Performance These systems must reliably Tunable Implement configurable sensitivity detect concerning patterns Detection settings with regular calibration while maintaining Thresholds based on operational experience manageable false positive and changing threat patterns. rates. 3.3.1 KEY TAKEAWAYS FOR BEHAVIOR OBSERVATION SYSTEMS • Clear definitions of "normal" behavior must be established with consideration for legitimate variations, • Privacy protections and appropriate notification are essential for ethical implementation, and • All alerts require human investigation to verify concerns before any actions are taken. 15 3.4 IMPAIRMENT DETECTION SYSTEMS Impairment detection systems are a specialized case of anomaly detection where the system Key Benefits leverages AI to identify potential fitness-for-duty ✓ Objective measurement of concerns by analyzing physiological and impairment indicators behavioral indicators of fatigue, intoxication, or ✓ Continuous or point-of-entry other impairment. These systems monitor screening capabilities visual cues, movement patterns, and other ✓ Non-invasive detection observable signals that may indicate a person methods is not fit to perform security- or safety-critical ✓ Earlier identification of fitness functions. As noted earlier, they are generally considered to be in service of an overall concerns assessment of personnel trustworthiness or reliability. In nuclear facilities, these technologies enhance existing fitness-for-duty programs by providing objective indicators to complement traditional testing methods. They serve as screening tools to identify potential impairment that may warrant further evaluation, helping maintain safety and security when personnel are performing sensitive functions. Table 7: Impairment Detection Systems Overview Technology Central Question Application Type Common Mathematical Application (What it's trying Algorithm Types Problem Type (How to do) (How it's Built) it Works) Fatigue Detection Is this individual displaying Computer Vision Neural Classification signs of fatigue that could Networks impact safety or security? Drug/Alcohol Is this individual displaying Multi-Modal Neural Classification Impairment signs of substance Networks Detection impairment? AI systems process multi-modal inputs to analyze and identify fatigue utilizing physiological and behavioral indicators. These indicators can include facial expressions, eye movement, breathing patterns, heart rate and other fatigue indicators to assess fatigue levels, reducing the risk of errors in security-sensitive roles. Two recent reviews of research in this area suggest that wearable and AI- enabled solutions provide a strong option for fatigue monitoring and detection [61, 62]. There is also a commercially-available dashcam that incorporates drowsiness detection, although such a product would need to be adapted for use in security applications in a large-scale facility [63]. AI-powered impairment detection systems analyze surveillance footage or other inputs (e.g., speech, physiological indicators) to identify potential signs of inebriation or impairment in personnel. In this application, an AI system generates alerts when it detects behavioral patterns consistent with impairment, supporting existing fitness-for-duty programs in nuclear facilities. There are a number of research papers that have explored machine learning algorithms to support detection of impairment, especially by alcohol, as well as a small start-up company that currently sells an AI-powered device for detection of driver impairment [64-67] . Before implementation of AI-powered impairment detection systems, organizations should first review fitness-for-duty policies and procedures to ensure that they are clear. Like other systems that inform or alter existing processes, they should discuss and determine how these systems will be integrated with existing approaches. There should be defined response procedures for detected impairment by the AI system that may different from those used for manual or human detection of 16 impairment. Just as with the broader fitness-for-duty program, before implementing AI for this use case, organizations should consider how they will protect private medical information and respect the dignity of their staff. As with other applications of AI technology, it is important to ensure that human personnel remain accountable for real-world decisions based on system outputs. Given the sensitivity of the data that these systems may be handling, it is especially important to ensure that it is sufficiently protected from attack or from misuse. Finally, because results of these AI systems may inform critical decisions, such as the continued employment of a member of facility staff, it is critical that the results are interpretable, accurate, and fair. Table 8 outlines practices at the organizational level to address these and other risks. Table 8: Responsible Implementation Checklist for Impairment Detection Systems Principle Key Considerations Managed through… Process Accountability Impairment allegations Multi-Step Verification Implement structured have significant verification protocols that consequences requiring require multiple forms of careful handling and evidence before making verification. impairment determinations or taking action. Fairness & Bias Systems may incorrectly Medical Review Process Establish a formal process for flag medical conditions, medical professionals to review disabilities, or other non- potential impairment impairment factors. indicators, considering legitimate medical explanations. Privacy & Data Impairment detection Medical Privacy Implement specialized privacy Protection involves collection of Framework protections aligned with sensitive health-related medical privacy standards, with information. strict separation between health data and personnel records. Security & Impairment allegations Independent Verification Require independent Abuse could be weaponized confirmation of all impairment Prevention against individuals indicators through different through system methods, with protection manipulation. against single-source allegations. Explainability & Personnel must Explainable/Transparent System must identify Interpretability understand the basis for Indicators signatures of impairment that impairment has elevated an individual to be determinations to ensure a concern and contributed to fair treatment. impairment determinations, with explanation appropriate for the individual. Safety Both false positives and Balanced Response Create tiered response negatives create safety Protocol procedures that appropriately risks through either balance the risks of both types removing needed of errors based on the specific personnel or allowing role and impairment indicators. impaired operations. Performance These systems must Performance Monitoring in Validate system performance perform consistently Variable Environments across varied operational across different conditions, lighting 17 individuals, environments, and times of day environments, and to ensure consistent detection conditions. capability. 3.4.1 KEY TAKEAWAYS FOR IMPAIRMENT DETECTION SYSTEMS • All potential impairment alerts require verification through established fitness-for-duty protocols, • Medical privacy considerations are essential due to the health-related nature of impairment data, and • Systems should be calibrated to consider medical conditions that may present similar indicators to impairment. 18 3.5 ACCESS CONTROL AND PATROLLING SYSTEMS Access control and patrolling systems employ AI to secure facility boundaries by Key Benefits authenticating personnel, screening for ✓ Enhanced perimeter and internal prohibited items, and conducting security monitoring autonomous security monitoring. These ✓ Consistent application of access systems control physical access through control policies biometric verification, detect unauthorized ✓ Extended surveillance coverage items through image analysis, and extend beyond fixed points surveillance capabilities through mobile ✓ Reduced personnel requirements platforms. In nuclear facilities, these for routine security functions technologies actively enforce security boundaries and expand monitoring coverage beyond fixed locations. They serve as front-line security systems making real-time decisions that directly impact facility security, often operating with limited human intervention in the immediate decision loop. Table 9: Access Control and Patrolling Systems Overview Technology Central Question Application Type Common Mathematical Application (What it's trying Algorithm Types Problem Type (How it to do) (How it's Built) Works) Facial Recognition Is this person Computer Vision Neural Networks Classification for Access Controls authorized to access this area? Automated Does this scan Computer Vision Neural Networks Classification Screening at contain prohibited Security items? Checkpoints Autonomous Is there unusual or Computer Vision Neural Networks Classification & Security Patrols concerning activity Clustering in this area? This application describes AI-powered biometric systems (including facial recognition and fingerprint verification) that control physical access to restricted areas within nuclear facilities by authenticating individual identities in real-time. Unlike background check applications of AI that could use facial recognition technology, this technology directly controls access decisions and requires immediate, accurate performance in operational conditions. AI-powered facial recognition for access control continues to be an active area of research, but there are also commercially available solutions that could be deployed in operations [68-71]. AI-powered systems to screen imagery at security checkpoints quickly assess items for prohibited materials with the goal of improving speed and accuracy over traditional screening methods. These systems assess images and data from security scans (e.g., X-ray) and determine the security status of each item. For example, TSA is exploring how it can use AI to enhance x-ray screening of luggage and personal items [72]. Similar technology could be deployed at security checkpoints at nuclear facilities to screen entrants for prohibited items. In this application of AI technology, uncrewed autonomous systems (UAS; ground vehicles or aerial vehicles) conduct security patrols using various sensors and computer vision to detect potential security concerns. These systems extend surveillance coverage beyond fixed cameras, operating 19 independently to identify and report anomalies while following predetermined patrol routes or responding to specific areas of interest. Although not a security application, the National Aeronautics and Space Administration has explored the use of UAS for fire detection and mapping applications [73, 74]. New York has also explored the use of an autonomous security robot—created by the company Knightscope—for subway patrols, although the pilot of the system was ultimately ended in 2024 [75, 76]. Because these technologies have real-time decision-making implications, it is especially important that facilities assess their readiness for implementation. This includes the integration of these AI systems into their overall physical security plan, a robust communication infrastructure to address any system issues or alerts in real-time, and clear response protocols for when those issues arrive. It is also critical that organizations identify backup or redundant security measures before implementing AI in this real-time fashion. In addition, the immediate impact of these systems on facility security necessitates a clear focus on measures to address risks of personnel safety, security, and accountability, and requires an especially high standard for their performance. That performance should be tested both prior to procurement and during deployment. Additionally, these systems require continuous human oversight through monitoring of outputs at regular defined intervals (e.g., audits, performance monitoring) and through unannounced review and testing. Provides more detail on the measures organizations can use to address these concerns. Table 10: Responsible Implementation Checklist for Access Control and Patrolling systems Principle Key Considerations Managed Process through… Accountability These systems directly control Human Oversight Maintain continuous human security boundaries with minimal Framework oversight of system operations human intervention in with clear responsibility chains immediate decisions. and immediate intervention capabilities, including periodic testing and auditing. Fairness & Bias Access control systems may Equitable Testing Regularly test and calibrate exhibit varying performance systems to ensure consistent across demographic groups, performance across all potentially creating unequal demographic groups, and barriers. suspend usages if disparities are identified. Privacy & Data Security monitoring creates Privacy by Design Implement technical controls Protection significant privacy implications, that limit surveillance to security- especially for continuous or relevant areas and information, biometric surveillance. with clear notice to affected individuals. Security & These systems form critical Defense in Depth Deploy multiple, overlapping Abuse security boundaries that security controls with regular Prevention adversaries may specifically penetration testing and security target for compromise. assessments by qualified third Commented [SP1]: Should this, and other instances where parties. you mention adversarial use, be linked to the AANT project? Explainability & Security personnel must Decision The system must provide Interpretability understand system decisions to Traceability detailed logs of all security effectively respond to events decisions with clear indication of and troubleshoot issues. contributing factors, enabling reconstruction of decision processes. 20 Safety Autonomous security systems Comprehensive Implement multiple layers of could create physical safety Safety Controls safety features including physical hazards through unexpected barriers, movement restrictions, movements, access restrictions, collision avoidance, and or other actions. emergency shutdown capabilities. Performance These systems must maintain Environmental Conduct rigorous testing across high performance across varied Stress Testing different weather conditions, environmental conditions and lighting situations, operational scenarios. scenarios and failure modes to ensure consistent security performance. 3.5.1 KEY INSIGHTS FOR IMPAIRMENT DETECTION SYSTEMS • Robust backup authentication methods must be maintained for system failures or performance limitations • Regular testing across environmental conditions is essential to ensure consistent security effectiveness • Human oversight capabilities must be maintained despite the autonomous nature of these systems 3.6 NUCLEAR MATERIAL ACCOUNTING AND CONTROL SYSTEMS AI-powered nuclear material accounting and control (NMAC) systems use AI to detect deviations in nuclear materials accounting data that may indicate theft of nuclear material. The anomaly detection system looks at nuclear accounting data and alerts when patterns of protracted or immediate theft occur. NMAC anomaly detection models are trained on either real or realistic but synthetic data. For example, the IAEA uses AI/ML models to create synthetic nuclear material accounting and verification data, which are then provided to external researchers to use in developing evaluation methodologies or internal training [77]. In another example, researchers experimented with clustering algorithms and graph analytics to support monitoring accounting of on-site material [78]. Especially for databases already stored in a digital format, AI presents potential for efficiency gains above and beyond traditional approaches [79]. Table 11: Nuclear Material Accounting and Control Systems Overview Technology Central Question Application Common Algorithm Mathematical Application Type (What it's Types (How it's Built) Problem Type trying to do) (How it Works) What are the normal patterns of material Material Movement movement at this Dimension Principal Component Characterization facility? Reduction Analysis Clustering Do current material transactions deviate Neural Networks, Transaction from expected Anomaly Random Forest, Classification & Anomaly Detection patterns? Detection Clustering Clustering How can realistic Generative training scenarios be Adversarial Synthetic Data created without Generative Networks, Variational Generation sensitive data? Models Autoencoders Generation 21 Dimension reduction approaches, such as principal or independent components analysis, have applications to NMAC in that they can help to characterize different transactions that are recorded in material databases [78]. This technique helps to summarize the types of transactions that occur, and can provide a clearer picture of how material at a facility is typically being used. In doing so, the results can serve as a baseline for understanding unusual or atypical behaviors that may be indicative of theft or diversion. Results can also help to provide resources in a more directed or thoughtful manner by discovering how material is most often used, potentially providing greater coordination. From a security perspective, however, results are most valuable as a basis for future anomaly detection. AI shows promise as a potential tool for detecting anomalous material transactions or material movement in material accounting databases [78, 79]. By applying AI algorithms to existing databases and continually monitoring transactions, users will be able to more readily identify anomalous transactions that may indicate inappropriate use, theft, or diversion. This can reduce the burden on security personnel to identify potential subtle indications of concern in material control data. This is particularly of benefit given that protracted theft may be challenging to identify in large databases, making AI a potentially valuable tool. Training personnel on identifying indications of material theft or diversion can be challenging given the paucity of real-world data available. Recently, the IAEA has explored the use of ML algorithms to create more realistic simulations to help support training of personnel [77]. This allows for the creation of more complex scenarios that better reflect realistic problem sets, and can facilitate information sharing with others for greater collaboration in the NMAC domain. To implement AI systems for nuclear material accounting and control (NMAC), organizations must first maintain a digital nuclear material accounting system with high-quality historical data, as these AI applications depend on accurate baseline patterns to effectively detect anomalies in material transactions and movements. Given the serious security implications of detected anomalies, prior to implementation, organizations must develop clear response protocols that define specific investigation steps, assign responsibilities for follow-up actions, and establish timelines for resolving flagged concerns. The sensitive nature of nuclear material data also requires robust protections to ensure the integrity and security of the data. Finally, successful implementation requires personnel who are trained in both traditional nuclear material accounting principles and modern data interpretation techniques, ensuring they can effectively evaluate AI system outputs within the proper operational and regulatory context. Once the decision to implement an AI system is made, organizations must establish processes for documenting system decisions, maintain multiple layers of security protections for the NMAC system itself, and employ strict role-based access control and audit logging to prevent system misuse or abuse. In addition, systems should be tested on a regular basis to ensure that they detect a variety of diversion methods. The system should also provide information about what triggered alerts in order to help maintain human accountability over NMAC decisions. Table 12 provides greater detail on processes that can help to address these risks. Table 12: Responsible Implementation Checklist for NMAC Systems Principle Key Considerations Managed through… Process Accountability NMAC decisions have Formal Alert Response Establish clear protocols significant security and Framework defining responsibilities for regulatory implications investigating and responding to 22 requiring clear system alerts, with responsibility structures. comprehensive documentation requirements. Fairness & Bias System may incorrectly Operational Context Incorporate operational context flag legitimate operational Integration into alert evaluation, ensuring variations as suspicious, normal process variations are potentially focusing properly distinguished from attention inappropriately. genuine anomalies. Privacy & Data While focused on material Focused Data Usage Implement strict controls Protection rather than personnel, ensuring NMAC data is used system may indirectly exclusively for material reveal information about accounting purposes, with individual actions. separation from personnel evaluation systems. Security & System could be targeted System Integrity Controls Deploy multiple layers of Abuse by adversaries seeking to security protections for the Prevention mask diversion activities NMAC system itself, with or create false alarms. rigorous access controls and comprehensive audit logging with regular human reviews of those logs. Explainability & Personnel must Explainable/Transparent Ensure system provides specific Interpretability understand specific Indicators information about which factors triggering alerts to patterns triggered alerts, with effectively investigate clear visualization of relevant potential diversions. data to support investigation. Safety Over-reliance on Alternative Detection Maintain traditional accounting automated systems could Methods methods alongside automated create vulnerabilities systems, with regular exercises through missed diversions to validate detection or distraction via false capabilities through multiple alarms. approaches. Performance System must reliably Diversion Scenario Regularly test system using detect both abrupt and Testing realistic diversion scenarios protracted diversion including various methods, scenarios while quantities, and timeframes to minimizing false alarms. ensure comprehensive detection capabilities. 3.6.1 KEY TAKEAWAYS FOR ACCESS CONTROL AND PATROLLING SYSTEMS • Robust backup authentication methods must be maintained for system failures or performance limitations, • Regular testing across environmental conditions is essential to ensure consistent security effectiveness, and • Human oversight capabilities must be maintained despite the autonomous nature of these systems. 23 3.7 ENGAGEMENT MATERIALS AND FACT SHEETS The information presented in this report is also summarized in a set of engagement materials designed for use with international partners. The engagement materials are intended to be used either during a standalone engagement of approximately one hour in length or as part of a larger engagement discussing insider threats and ITM. They are most suited for peer or near-peer partners as part of a mutual technical exchange, as this is an emerging technology with limited regulation or standards. Possible partners include the United Kingdom, Australia, or Belgium. Additionally, we have created a set of fact sheets for each category of AI technology that can be left as read-behind materials for partners to consider their readiness to implement AI technology in this manner. They contain a brief version of some of the content presented here, including a short description of the technologies, benefits, prerequisites, important notes, and the “Responsible Implementation Checklists” shown in each section of this report. These fact sheets are presented in Appendix A. They should ideally be presented as part of the larger engagement to ensure that partners sufficiently understand the context and caveats regarding the material: they are intended primarily as reminders rather than comprehensive documents representing every possible risk of an AI system for each application. 24 CHAPTER 4 4. CONCLUSION AI applications offer significant potential to enhance nuclear security ITM programs by providing capabilities that extend beyond traditional human-centered approaches. These technologies can process vast amounts of data continuously, detect subtle patterns across multiple systems, and provide consistent application of security criteria—advantages that are particularly valuable given the complex and often subtle nature of insider threats. However, the implementation of AI in nuclear security contexts also introduces new risks that must be carefully managed to ensure these systems enhance rather than undermine overall security posture. This framework provides organizations with a structured approach to evaluating and implementing AI technologies for ITM while maintaining appropriate human oversight and addressing the ethical and operational challenges inherent in these applications. The six functional categories identified in this analysis—from identity verification through nuclear material accounting—represent different approaches to addressing insider threat challenges, each with distinct operational requirements and organizational considerations. Organizations must carefully assess their specific needs, capabilities, and operational context when considering which AI applications align with their security objectives. Organizations must also adopt tailored approaches for each category of AI application they choose to pursue. Identity and document verification systems, for example, build upon well-established commercial technologies with clear performance metrics and relatively well-defined implementation pathways. Behavior observation systems, in contrast, require sophisticated organizational infrastructure, extensive policy development, and ongoing validation efforts. Nuclear material accounting applications leverage established mathematical approaches while requiring specialized expertise and regulatory compliance considerations. Each category demands careful planning and preparation to support successful deployment. Critically, all AI applications for ITM should be implemented as decision support tools that enhance human capabilities rather than replace human judgment. The high-consequence nature of nuclear security decisions, combined with the current limitations of AI technology and the sparse data available for insider threat scenarios, makes human oversight essential for responsible implementation. Organizations must establish clear protocols for how AI system outputs will be used, verified, and acted upon, ensuring that human operators understand both the capabilities and limitations of these technologies. The risk management approach outlined in this framework, based on the NIST AI Risk Management Framework principles, provides a method for addressing the challenges of AI deployment in nuclear security contexts. By systematically considering accountability, fairness, privacy, security, explainability, safety, and performance considerations, organizations can work to maximize the benefits of AI technologies while minimizing potential negative consequences. However, this requires ongoing commitment to monitoring system performance, updating procedures as technology evolves, and maintaining organizational capabilities for both AI-enabled and traditional security approaches. Looking forward, organizations should recognize that responsible AI implementation for ITM is not a one-time effort but an ongoing process that requires continuous evaluation and adaptation. As AI technologies continue to evolve and new applications emerge, the frameworks and procedures 25 established today must be regularly updated to reflect changing capabilities, threats, and regulatory requirements. Organizations that invest in developing strong foundational capabilities for AI risk management today will be better positioned to adapt to future technological developments while maintaining the highest standards of nuclear security. 4.1 IMPLEMENTATION CONSIDERATIONS AND ORGANIZATIONAL READINESS Before pursuing AI applications for ITM, organizations must carefully assess their readiness across multiple dimensions. Technical readiness involves not only the availability of appropriate data and IT infrastructure, but also the organizational capacity to integrate AI systems with existing security processes without creating vulnerabilities or operational disruptions. This includes ensuring that personnel responsible for system operation and oversight have appropriate training in both the technical aspects of AI systems and their specific applications to nuclear security contexts. Organizational readiness also encompasses policy and procedural considerations. Many AI applications for ITM involve sensitive areas including personnel privacy, employment decisions, and regulatory compliance. Organizations must ensure they have appropriate legal and policy frameworks in place before implementation, including clear guidelines for data handling, system use limitations, and response protocols for different types of AI-generated alerts or recommendations. This may require coordination across multiple organizational functions including security, human resources, legal, and IT departments. Different AI applications present varying operational requirements, and organizations must carefully consider whether their current capabilities are sufficient for their intended implementations. Applications such as comprehensive behavioral monitoring or trustworthiness assessments may require more extensive oversight mechanisms, validation procedures, and backup systems than applications like document verification or material accounting anomaly detection. Understanding these requirements early in the planning process helps ensure successful implementation. Finally, organizations should consider their long-term strategic approach to AI adoption for ITM. Rather than implementing individual systems in isolation, organizations may benefit from developing comprehensive roadmaps that sequence different AI applications based on their organizational readiness and potential security benefits. This strategic approach can help ensure that foundational capabilities developed for initial AI implementations can support additional applications over time, maximizing return on investment while maintaining security effectiveness. By following the framework presented in this analysis and carefully considering their organizational readiness across these dimensions, nuclear security organizations can work toward responsible implementation of AI technologies that enhance their ITM capabilities while maintaining the highest standards of safety, security, and ethical operation. 4.2 CAVEATS AND LIMITATIONS Critically, the risks and considerations presented in this framework are not comprehensive and should not be considered exhaustive. Although we have applied established risk management principles from the NIST AI Risk Management Framework and drawn on a recent literature review, the specific operational contexts, regulatory environments, and organizational characteristics of individual nuclear facilities may introduce additional risks or considerations not addressed in this analysis. Organizations should view this framework as a foundation for their own risk assessment processes rather than a complete catalog of all potential concerns. Each organization must conduct its own thorough of how AI applications may interact with their specific operational procedures, 26 security measures, and regulatory requirements to identify additional risks or mitigation strategies that may be necessary for their unique circumstances. Second, this document will likely become outdated relatively quickly due to the rapid pace of advancement in AI capabilities and applications. The field of AI is evolving at an unprecedented rate, with new algorithms, approaches, and commercial products emerging regularly. What represents cutting-edge or experimental technology today may become standard practice within months or years, while entirely new categories of AI applications may emerge that are not currently anticipated. Similarly, regulatory frameworks, industry standards, and best practices for AI deployment in high- consequence domains are still developing and may change significantly as governments and organizations gain more experience with these technologies. Organizations should therefore treat this analysis as a snapshot of current capabilities and considerations rather than a definitive long- term guide, and should plan to regularly update their AI strategies and risk assessments as the technology landscape continues to evolve. Third, this analysis focuses specifically on core AI capabilities that directly support ITM functions as defined by the IAEA framework. However, there are many other AI systems and capabilities that may be relevant to ITM in more ancillary or indirect ways that are not reviewed in this document. For example, AI applications for general cybersecurity monitoring, administrative process automation, or facility management may have implications for insider threat programs even if they are not primarily designed for that purpose. Similarly, advances in AI technologies for areas such as social media analysis, communication monitoring, or predictive analytics may create new opportunities or challenges for ITM that are not fully explored here. Organizations should consider how their broader adoption of AI technologies across different operational areas may interact with their ITM strategies and may require additional analysis beyond what is provided in this framework. Additionally, the commercial AI landscape is highly dynamic, with new vendors, products, and capabilities entering the market regularly while others may be discontinued or significantly modified. The specific commercial examples referenced in this analysis should be understood as illustrative of current capabilities rather than endorsements of particular vendors or products. Organizations should conduct their own market research and vendor evaluation processes when considering specific AI implementations. Finally, while this framework addresses many of the key considerations for responsible AI implementation in nuclear security contexts, it cannot substitute for the specialized expertise, regulatory guidance, and organizational knowledge required for successful deployment. Organizations should ensure they have access to appropriate technical, legal, and policy expertise throughout their AI implementation processes and should work closely with relevant regulatory bodies to ensure compliance with all applicable requirements. The high-consequence nature of nuclear security applications demands careful, methodical approaches to AI adoption that prioritize safety and security above other considerations, and organizations should be prepared to invest the time and resources necessary to implement these technologies responsibly. 27 5. REFERENCES [1] S. Yuan and X. Wu, "Deep learning for insider threat detection: Review, challenges and opportunities," Computers & Security, vol. 104, p. 102221, 2021/05/01/ 2021, doi: https://doi.org/10.1016/j.cose.2021.102221. [2] A. D. Williams et al., "A New Approach to Insider Threat Detection & Mitigation for Critical Infrastructure," in Workshop on Research for Insider Threats, Austin, TX, 4 - 8 December 2023. [3] A. D. Williams, S. N. Abbott, N. Shoman, and W. S. Charlton, "Results From Invoking Artificial Neural Networks to Measure Insider Threat Detection & Mitigation," Digital Threats, vol. 3, no. 1, p. Article 3, 2021, doi: 10.1145/3457909. [4] F. R. Alzaabi and A. Mehmood, "A Review of Recent Advances, Challenges, and Opportunities in Malicious Insider Threat Detection Using Machine Learning Methods," IEEE Access, vol. 12, pp. 30907-30927, 2024, doi: 10.1109/ACCESS.2024.3369906. [5] E. Yilmaz and O. Can, "Unveiling Shadows: Harnessing Artificial Intelligence for Insider Threat Detection," Engineering, Technology & Applied Science Research, vol. 14, no. 2, pp. 13341- 13346, 04/02 2024, doi: 10.48084/etasr.6911. [6] Z. Buçinca, M. B. Malaya, and K. Z. Gajos, "To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making," Proceedings of ACM on Human- Computer Interaction, vol. 5, no. CSCW1, p. Article 188, 2021, doi: 10.1145/3449287. [7] IBM. "AI Risk Atlas." https://www.ibm.com/docs/en/watsonx/saas?topic=ai-risk-atlas (accessed 14 April, 2025). [8] P. Slattery et al., "The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks From Artificial Intelligence," p. arXiv:2408.12622doi: 10.48550/arXiv.2408.12622. [9] NIST, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)," 2023. [10] A. Brenneis, "Assessing dual use risks in AI research: necessity, challenges and mitigation strategies," Research Ethics, vol. 21, no. 2, pp. 302-330, 2025, doi: 10.1177/17470161241267782. [11] U. Anwar et al., "Foundational Challenges in Assuring Alignment and Safety of Large Language Models," doi: 10.48550/arXiv.2404.09932. [12] A. Wasil, J. Clymer, D. Krueger, E. Dardaman, S. Campos, and E. Murphy, "Affirmative Safety: An Approach to Risk Management for Advanced Ai," Available at SSRN 4806274, 2024. [13] A. A. Mughal, "Artificial Intelligence in Information Security: Exploring the Advantages, Challenges, and Future Directions," Journal of Artificial Intelligence and Machine Learning in Management, vol. 2, no. 1, pp. 22-34, 01/20 2018. [Online]. Available: https://journals.sagescience.org/index.php/jamm/article/view/51. [14] C. Huang, Z. Zhang, B. Mao, and X. Yao, "An Overview of Artificial Intelligence Ethics," IEEE Transactions on Artificial Intelligence, vol. 4, no. 4, pp. 799-819, 2023, doi: 10.1109/TAI.2022.3194503. [15] J. R. Biden, "Executive order on the safe, secure, and trustworthy development and use of artificial intelligence," 2023. [16] IAEA, Preventive and Protective Measures Against Insider Threats. Vienna, Austria: International Atomic Energy Agency, 2020. [17] E. Shaw, K. G. Ruby, and J. M. Post, "The insider threat to information systems," Security Awareness Bulletin, vol. 2, no. 98, 1998. [18] M. Bunn and S. D. Sagan, "Introduction: Inside the Insider Threat," in Insider Threats: Cornell University Press, 2017. [19] M. Bunn and S. D. Sagan, "A Worst Practices Guide to Insider Threats," in Insider Threats: Cornell University Press, 2017. 28 [20] F. L. Greitzer, L. J. Kangas, C. F. Noonan, and A. C. Dalton, "Identifying at-risk employees: A behavioral model for predicting potential insider threats," Pacific Northwest National Lab. (PNNL), Richland, WA (United States), United States, 2010. [Online]. Available: https://www.osti.gov/biblio/1000159 https://www.osti.gov/servlets/purl/1000159 [21] F. Greitzer, J. Purl, Y. M. Leong, and D. S. Becker, "Sofit: Sociotechnical and organizational factors for insider threat," in 2018 IEEE Security and Privacy Workshops (SPW), 2018: IEEE, pp. 197-206. [22] V. Koutsouvelis, S. Shiaeles, B. Ghita, and G. Bendiab, "Detection of Insider Threats using Artificial Intelligence and Visualisation," in 2020 6th IEEE Conference on Network Softwarization (NetSoft), 29 June-3 July 2020 2020, pp. 437-443, doi: 10.1109/NetSoft48620.2020.9165337. [23] Canadian Nuclear Safety Commission UK Office for Nuclear Regulation US Nuclear Regulatory Commission. (2024). ML24241A252, Considerations for Developing Artificial intelligence Systems in Nuclear Applications. [Online] Available: https://www.nrc.gov/docs/ML2424/ML24241A252.pdf [24] U.S. Nuclear Regulatory Commission. (2024). ML24290A059, Regulatory Framework Gap Assessment for the use of Artificial Intelligence in Nuclear Applications. [Online] Available: https://www.nrc.gov/docs/ML2429/ML24290A059.pdf [25] E. Glikson and A. W. Woolley, "Human Trust in Artificial Intelligence: Review of Empirical Research," Academy of Management Annals, vol. 14, no. 2, pp. 627-660, 2020, doi: 10.5465/annals.2018.0057. [26] K. E. Henry et al., "Human–machine teaming is key to AI adoption: clinicians’ experiences with a deployed machine learning system," npj Digital Medicine, vol. 5, no. 1, p. 97, 2022/07/21 2022, doi: 10.1038/s41746-022-00597-7. [27] S. S. Y. Kim, Q. V. Liao, M. Vorvoreanu, S. Ballard, and J. W. Vaughan, ""I'm Not Sure, But...": Examining the Impact of Large Language Models' Uncertainty Expression on User Reliance and Trust," presented at the Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, Rio de Janeiro, Brazil, 2024. [Online]. Available: https://doi.org/10.1145/3630106.3658941. [28] Z. Buçinca, M. B. Malaya, and K. Z. Gajos, "To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making," Proc. ACM Hum.-Comput. Interact., vol. 5, no. CSCW1, p. Article 188, 2021, doi: 10.1145/3449287. [29] S. Casper et al., "Black-Box Access is Insufficient for Rigorous AI Audits," doi: 10.48550/arXiv.2401.14446. [30] Y. Zhang, Q. V. Liao, and R. K. E. Bellamy, "Effect of confidence and explanation on accuracy and trust calibration in AI-assisted decision making," presented at the Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, Barcelona, Spain, 2020. [Online]. Available: https://doi.org/10.1145/3351095.3372852. [31] S. Ling, Y. Zhang, and N. Du, "Impacts of AI-Generated Confidence and Explanations on Task Performance and Trust in Human-autonomy Teaming," Proceedings of the Human Factors and Ergonomics Society Annual Meeting, vol. 0, no. 0, p. 21695067231195002, 2023, doi: 10.1177/21695067231195002. [32] U. Bhatt et al., "Uncertainty as a Form of Transparency: Measuring, Communicating, and Using Uncertainty," presented at the Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, Virtual Event, USA, 2021. [Online]. Available: https://doi.org/10.1145/3461702.3462571. 29 [33] C. Janiesch, P. Zschech, and K. Heinrich, "Machine learning and deep learning," Electronic Markets, vol. 31, no. 3, pp. 685-695, 2021/09/01 2021, doi: 10.1007/s12525-021-00475- 2. [34] H.-P. Lee, Y.-J. Yang, T. Serban von Davier, J. Forlizzi, and S. Das, "Deepfakes, Phrenology, Surveillance, and More! A Taxonomy of AI Privacy Risks," p. arXiv:2310.07879doi: 10.48550/arXiv.2310.07879. [35] K. Wach et al., "The dark side of generative artificial intelligence: A critical analysis of controversies and risks of ChatGPT," (in English), Entrepreneurial Business and Economics Review, vol. 11, no. 2, pp. 7-30, 2023, doi: https://doi.org/10.15678/EBER.2023.110201. [36] H. A. Inan et al., "Training Data Leakage Analysis in Language Models," p. arXiv:2101.05405doi: 10.48550/arXiv.2101.05405. [37] O.-M. C. Osazuwa and M. O. Musa, "The Expanding Attack Surface: Securing AI and Machine Learning Systems in Security Operations," International Journal of Innovative Science and Research Technology, vol. 9, pp. 2498-2505, 2024. [38] A. T. Olutimehin, A. J. Ajayi, O. C. Metibemu, A. Y. Balogun, T. O. Oladoyinbo, and O. O. Olaniyi, "Adversarial Threats to AI-Driven Systems: Exploring the Attack Surface of Machine Learning Models and Countermeasures," (in English), Journal of Engineering Research and Reports, vol. 27, no. 2, pp. 341-362, 2025-02-13 2025, doi: 10.9734/jerr/2025/v27i21413. [39] D. Nigenda et al., "Amazon SageMaker Model Monitor: A System for Real-Time Insights into Deployed Machine Learning Models," presented at the Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington DC, USA, 2022. [Online]. Available: https://doi.org/10.1145/3534678.3539145. [40] D. Leslie, "Understanding artificial intelligence ethics and safety," p. arXiv:1906.05684doi: 10.48550/arXiv.1906.05684. [41] C. AI. "Clearview AI." https://www.clearview.ai/ (accessed 2 May, 2025). [42] L. R. Solutions. "LexID® and Scalable Automated Linking Technology (SALT)." https://risk.lexisnexis.com/products/lexid-and-salt-for-government (accessed 12 February, 2025). [43] Veriff. "Identity Verification." https://www.veriff.com/product/identity-verification (accessed 2 May, 2025). [44] checkr. "AI Powered Core." https://checkr.com/our-technology/ai-powered (accessed 2 May, 2025). [45] L. A. Garcia-Segura, "The role of artificial intelligence in preventing corporate crime," Journal of Economic Criminology, vol. 5, p. 100091, 2024/09/01/ 2024, doi: https://doi.org/10.1016/j.jeconc.2024.100091. [46] T. Shahana, V. Lavanya, and A. R. Bhat, "State of the art in financial statement fraud detection: A systematic review," Technological Forecasting and Social Change, vol. 192, p. 122527, 2023/07/01/ 2023, doi: https://doi.org/10.1016/j.techfore.2023.122527. [47] (2024). Managing Artificial Intelligence-Specific Cybersecurity Risks in the Financial Services Sector. [Online] Available: /https://home.treasury.gov/system/files/136/Managing-Artificial- Intelligence-Specific-Cybersecurity-Risks-In-The-Financial-Services-Sector.pdf [48] Airbnb. "Aircover for Hosts." https://www.airbnb.com/aircover-for-hosts (accessed 2 May, 2025). [49] X. Dastile, T. Celik, and M. Potsane, "Statistical and machine learning models in credit scoring: A systematic literature survey," Applied Soft Computing, vol. 91, p. 106263, 2020/06/01/ 2020, doi: https://doi.org/10.1016/j.asoc.2020.106263. [50] (2019). WP 18-15, The Roles of Alternative Data and Machine Learning in Fintech Lending: Evidence from the Lending Club Consumer Platform. [Online] Available: https://www.philadelphiafed.org/-/media/FRBP/Assets/working-papers/2018/wp18- 15R.pdf?sc_lang=en 30 [51] (2018). Financial Technology: Agencies Should Provide Clarification on Lenders' Use of Alternative Data. [Online] Available: https://www.gao.gov/assets/d19111.pdf [52] D. C. a. S. Agency. "Continuous Vetting." https://www.dcsa.mil/Personnel- Security/Continuous-Vetting/ (accessed 2 May, 2025). [53] N. Carignan, "From Hype to Reality: How AI is Transforming Cybersecurity Practices," in Darktrace, ed. https://www.darktrace.com/blog/how-ai-is-transforming-cybersecurity- practices, 2025. [54] E. Treviño et al., "Autonomous anomaly detection of proliferation in the AGN-201 nuclear reactor digital twin," Annals of Nuclear Energy, vol. 211, p. 110990, 2025/02/01/ 2025, doi: https://doi.org/10.1016/j.anucene.2024.110990. [55] A. D. Williams, S. N. Abbott, N. Shoman, and W. S. Charlton, "Results From Invoking Artificial Neural Networks to Measure Insider Threat Detection & Mitigation," Digital Threats, vol. 3, no. 1, 2021, doi: 10.1145/3457909. [56] Verkada. "AI Video Analytics." https://info.verkada.com/surveillance-features/ai-video- analytics/ (accessed 2 May, 2025). [57] E. Perez, "Using Video Content Analytics to Track Trends, Identify Anomalies and Accelerate Real-Time Response," in AI and Video Analytics Blog vol. 2025, ed: Briefcam, 2020. [58] actuate. "Real-Time AI Video Analytics." https://www.briefcam.com/resources/blog/using- video-content-analytics-to-track-trends-identify-anomalies-and-accelerate-real-time-response/ (accessed 2 May, 2025). [59] IARPA. "DIVA: Deep Intermodal Video Analytics." https://www.iarpa.gov/research- programs/diva (accessed 2 May, 2025). [60] L. Stanham. "What is AI-Powered Behavioral Analysis in Cybersecurity." Crowdstrike. https://www.crowdstrike.com/en-us/cybersecurity-101/artificial-intelligence/ai-powered- behavioral-analysis/ (accessed 5 May, 2025). [61] R. Hooda, V. Joshi, and M. Shah, "A comprehensive review of approaches to detect fatigue using machine learning techniques," Chronic Diseases and Translational Medicine, 2021/08/25/ 2021, doi: https://doi.org/10.1016/j.cdtm.2021.07.002. [62] K. Kakhi, S. K. Jagatheesaperumal, A. Khosravi, R. Alizadehsani, and U. Rajendra Acharya, "Fatigue Monitoring Using Wearables and AI: Trends, Challenges, and Future Opportunities," p. arXiv:2412.16847doi: 10.48550/arXiv.2412.16847. [63] samsara. "Stay Awake, Stay Safe: Real-time Alerts for Drowsy Drivers." https://www.samsara.com/products/safety/drowsiness-detection (accessed 2 May, 2025). [64] A. A. Bonela, Z. He, A. Nibali, T. Norman, P. G. Miller, and E. Kuntsche, "Audio-based Deep Learning Algorithm to Identify Alcohol Inebriation (ADLAIA)," Alcohol, vol. 109, pp. 49-54, 2023/06/01/ 2023, doi: https://doi.org/10.1016/j.alcohol.2022.12.002. [65] F. Amato, V. Cesarini, G. Olmo, G. Saggio, and G. Costantini, "Beyond breathalyzers: AI- powered speech analysis for alcohol intoxication detection," Expert Systems with Applications, vol. 262, p. 125656, 2025/03/01/ 2025, doi: https://doi.org/10.1016/j.eswa.2024.125656. [66] P. Laptev, V. Demareva, S. Litovkin, E. Kostuchenko, and A. Shelupanov, "Machine learning- based detection of alcohol intoxication through speech analysis: a comparative study of AI models," The European Physical Journal Special Topics, 2025/02/27 2025, doi: 10.1140/epjs/s11734-025-01508-z. [67] SoberRide. "SoberRide." https://www.sober-ride.com/ (accessed 2 May, 2025). [68] H. Lee, S.-H. Park, J.-H. Yoo, S.-H. Jung, and J.-H. Huh, "Face Recognition at a Distance for a Stand-Alone Access Control System," Sensors, vol. 20, no. 3, doi: 10.3390/s20030785. [69] R. Rameswari, S. Naveen Kumar, M. Abishek Aananth, and C. Deepak, "Automated access control system using face recognition," Materials Today: Proceedings, vol. 45, pp. 1251-1256, 2021/01/01/ 2021, doi: https://doi.org/10.1016/j.matpr.2020.04.664. 31 [70] alcatraz. "Meet the Rock: Next Level Access Control." https://alcatraz.ai/the-rock (accessed 2 May, 2025). [71] I. Group. "Facial Recognition Access Control." https://www.idemia.com/facial-recognition- access-control (accessed 2 May, 2025). [72] E. Graham, "TSA is Looking at Using AI to Improve Security Screening Processes," 8 May 2024. [Online]. Available: https://www.nextgov.com/artificial-intelligence/2024/05/how-tsa- looking-using-ai-improve-security-screening-processes/396398/ [73] NASA. "AI-Enabled Drone Swarms for Fire Detection, Mapping, and Modeling." https://esto.nasa.gov/firetech/ai-enabled-drone-swarms-for-fire-detection-mapping-and- modeling/ (accessed 2 May, 2025). [74] S. P. H. Boroujeni et al., "A comprehensive survey of research towards AI-enabled unmanned aerial systems in pre-, active-, and post-wildfire management," Information Fusion, vol. 108, p. 102369, 2024/08/01/ 2024, doi: https://doi.org/10.1016/j.inffus.2024.102369. [75] Knightscope. "Why Choose the Knightscope K5 Autonomous Security Robot?" https://knightscope.com/products/k5 (accessed 2 May, 2025). [76] D. M. Rubinstein, Hurubie, "Goodbye for Now to the Robot That (Sort Of) Patrolled New York’s Subway," in The New York Times, ed. New York, 2024. [77] P. S. Schneeweiss, T.; Baude, S., "The IAEA’s innovative approach to address the challenges in the collection and analysis of safeguards relevant information.," presented at the Institute of Nuclear Materials Management (INMM) Annual Meeting, Portland, OR, 2024. [Online]. Available: https://cdn.fourwaves.com/static/media/formdata/b12613f9-28b7-4568-a069- 5c7cfeff3c8e/8d5ee6c8-9afc-4bf0-812c-2f57cf07b9d9.pdf. [78] A. Drescher, M. Adams, S. Stewart, K. J. Dayman, L. G. Worrall, and G. Westphal, "Machine Learning Approaches for Nuclear Material Accounting Data from Irradiation and Reprocessing," presented at the Conference: 61st INMM Annual Meeting - Virtual Meeting (no city), Tennessee, United States of America - 7/12/2020 4:00:00 PM-7/16/2020 4:00:00 PM, United States, 2020. [Online]. Available: https://www.osti.gov/biblio/1649128 https://www.osti.gov/servlets/purl/1649128. [79] (2022). Artificial Intelligence for Accelerating Nuclear Applications, Science and Technology. [Online] Available: https://www.iaea.org/publications/15198/artificial-intelligence-for- accelerating-nuclear-applications-science-and-technology 32