National Institute of Standards and Technology AI 100-2e2025, Adversarial Machine Learning, examines the security risks posed by malicious actors who intentionally manipulate machine learning systems and outlines strategies to strengthen their resilience. The report explains how adversarial attacks can occur during different phases of the AI lifecycle, including data poisoning during training, model evasion through carefully crafted inputs, model extraction, and inference-time manipulation. It emphasizes that AI systems introduce new attack surfaces beyond traditional cybersecurity threats, requiring specialized risk assessment, testing, and monitoring approaches. The publication promotes secure-by-design principles, robust evaluation techniques, red-teaming, and continuous monitoring to detect and mitigate adversarial behaviors. Ultimately, NIST AI 100-2e2025 reinforces the need to integrate AI security into broader risk management and governance frameworks, ensuring machine learning systems remain reliable, trustworthy, and resilient in adversarial environments.

Organization for Economic Co-operation and Development Framework for Classifying AI Systems provides a structured method for categorizing AI systems based on their context, capabilities, and potential impact. Rather than ranking systems as simply “high” or “low” risk, the framework analyzes AI across multiple dimensions, including the system’s purpose, the data it uses, its degree of autonomy, the human involvement in decision-making, and the severity and scale of potential outcomes. It emphasizes lifecycle thinking—from design and development to deployment and monitoring—while encouraging policymakers and organizations to consider how AI systems interact with social, legal, and economic environments. The framework complements the OECD AI Principles by offering practical guidance for understanding risk profiles, supporting regulatory clarity, and promoting responsible, transparent, and accountable AI governance across sectors and jurisdictions.

National Institute of Standards and Technology AI 100-5, A Plan for Global Engagement on AI Standards, outlines a strategic approach for advancing international collaboration in the development of AI standards. The publication emphasizes the importance of aligning technical standards, risk management practices, and governance frameworks across countries to promote innovation while safeguarding trust, safety, and human rights. It identifies key priorities such as strengthening U.S. participation in international standards bodies, fostering public–private partnerships, promoting interoperability, and supporting capacity-building efforts worldwide. The report underscores that AI is inherently global, requiring coordinated engagement to address shared challenges like bias, security, transparency, and accountability. Ultimately, NIST AI 100-5 seeks to position standards development as a cornerstone of responsible AI adoption, ensuring that emerging technologies are governed through consistent, inclusive, and globally informed practices.

National Institute of Standards and Technology AI 100-3, The Language of Trustworthy AI, establishes a shared vocabulary to support consistent understanding and implementation of trustworthy artificial intelligence. The publication clarifies key concepts such as reliability, robustness, safety, security, resilience, privacy, fairness, accountability, transparency, and explainability, emphasizing that these characteristics are interconnected and context-dependent. By standardizing terminology, the report helps organizations, policymakers, and researchers communicate more precisely about AI risks, controls, and governance practices. It supports the broader AI Risk Management Framework by reducing ambiguity in how trustworthiness is defined and operationalized across sectors. Ultimately, NIST AI 100-3 promotes clearer communication, improved collaboration, and more effective governance, enabling stakeholders to design, evaluate, and deploy AI systems that align with societal values and organizational objectives.

National Institute of Standards and Technology Interagency Report 8367, Psychological Foundations of Explainability and Interpretability in AI, examines explainable AI through the lens of human cognition and decision-making. Rather than focusing solely on technical transparency, the report emphasizes how people understand, trust, and interact with AI systems. It explains that effective explanations must align with human mental models, account for cognitive limitations, and support users’ goals, context, and expertise levels. The publication distinguishes between interpretability (how well a system’s internal mechanics can be understood) and explainability (how well a system communicates its reasoning to humans), highlighting that explanations must be accurate, meaningful, and usable to improve trust calibration rather than blind trust. Ultimately, NISTIR 8367 argues that explainable AI is not purely a technical challenge but a socio-technical one, requiring interdisciplinary collaboration across AI development, psychology, and human-computer interaction to ensure AI systems are understandable, accountable, and appropriately trusted.

National Institute of Standards and Technology Interagency Report 8312, Four Principles of Explainable Artificial Intelligence, defines foundational guidance for designing AI systems that can be understood and trusted by users. The report outlines four core principles: AI systems should provide explanations for their outputs; those explanations should be meaningful and understandable to the intended audience; the explanations should accurately reflect how the system produced its results; and the system should operate only within the conditions for which it was designed and validated (knowledge limits). NISTIR 8312 emphasizes that explainability is context-dependent—what counts as a sufficient explanation varies across domains such as healthcare, finance, or criminal justice. Rather than prescribing a single technical method, the report promotes a risk-informed, human-centered approach that balances transparency, performance, and usability, helping organizations build AI systems that are more accountable, reliable, and aligned with stakeholder expectations.

This podcast outlines a structured approach to recognizing, measuring, and mitigating bias in artificial intelligence systems. The NIST publication distinguishes between statistical bias (systematic error in model outputs) and societal bias (inequities embedded in data and social systems), emphasizing that AI systems can both reflect and amplify existing disparities. It proposes a lifecycle-based framework that spans data collection, model development, evaluation, deployment, and monitoring, encouraging organizations to adopt transparency, documentation, testing, and stakeholder engagement practices. The report highlights the importance of clear definitions, contextual understanding of fairness, and measurable performance criteria, while acknowledging trade-offs among fairness metrics. Ultimately, SP 1270 aims to support the development of technical standards and governance practices that promote trustworthy, equitable, and accountable AI systems across sectors.

This podcast explores the core elements of NIST AI 600-1, a framework designed to help organizations identify, measure, and manage the unique and often amplified risks associated with generative AI systems. It highlights twelve central risk areas, including confabulation, dangerous or violent content, data privacy concerns, environmental impact, harmful and systemic bias, human–AI configuration challenges, information security threats, intellectual property risks, obscene or abusive outputs, and vulnerabilities across the AI value chain. The discussion also focuses on practical, cross-sector, and voluntary actions organizations can take to govern, map, measure, and manage these risks effectively. Special emphasis is placed on pre-deployment testing, content provenance mechanisms such as watermarking and tracking, and strong incident reporting processes to ensure accountability. Throughout, the podcast aligns these efforts with the AI Risk Management Framework’s four core functions—Govern, Map, Measure, and Manage—promoting greater trust, security, and safety in the deployment of generative AI technologies.

This podcast centers on the NIST AI Risk Management Framework (AI RMF), a voluntary standard designed to increase the trustworthiness of artificial intelligence through four key functions: govern, map, measure, and manage. While the core NIST documents establish foundational principles for mitigating sociotechnical harms—including specific risks like confabulation and bias in generative AI—supplementary research introduces a maturity model to help organizations operationalize these guidelines. This model provides a structured questionnaire and scoring system based on metrics such as robustness, coverage, and stakeholder diversity. By bridging the gap between high-level ethics and day-to-day practices, these resources offer a roadmap for evaluating an organization’s progress in managing complex AI threats. Ultimately, the collection emphasizes that responsible AI requires continuous monitoring, evidence-based accountability, and a deep understanding of the risks throughout the entire system lifecycle.

This podcast outlines the continuous monitoring phase of the Risk Management Framework (RMF), emphasizing the need for ongoing situational awareness in federal information systems. The provided text details how organizations must systematically track changes to technology, personnel, and operational environments to ensure that security and privacy controls remain effective over time. Key processes include performing security impact analyses, conducting regular audits, and utilizing automated tools like SCAP and SIEM for efficient data collection. The documentation also highlights the importance of reporting risk posture to authorizing officials to support near real-time authorization decisions. Finally, the guides address the end of the system life cycle, providing protocols for secure media sanitization and the formal decommissioning of information systems.