AI Governance & AI Security Frameworks Reference

AI Governance & AI Security Frameworks Reference

Reference of 30 AI governance, ethics, and security frameworks — covering regulatory requirements like the EU AI Act, technical standards like NIST AI RMF and ISO 42001, and adversarial threat frameworks like MITRE ATLAS and OWASP LLM Top 10.

• EU AI Act — risk-based AI regulation: prohibited practices, high-risk system requirements, and conformity assessments
• NIST AI RMF 1.0 — Govern, Map, Measure, and Manage functions for responsible AI risk management
• ISO 42001 — international AI management system standard (published December 2023)
• MITRE ATLAS — adversarial threat landscape for AI systems: attack techniques and mitigations
• OWASP LLM Top 10 — ten most critical security risks for large language model applications
• OECD AI Principles — internationally agreed values and guidelines for trustworthy AI

EU AI Act: Risk-Based AI Regulation and Compliance Requirements

The EU Artificial Intelligence Act, fully in force from August 2024 with phased obligations through 2027, is the world's first comprehensive AI regulation — applying a risk-based classification system that determines compliance obligations based on the intended use and potential harm of each AI system. Prohibited AI practices (unacceptable risk) were banned from February 2025, including social scoring, real-time biometric identification in public spaces for law enforcement (with exceptions), and manipulation of vulnerable groups. High-risk AI systems (employment decisions, credit scoring, biometric verification, critical infrastructure management) require conformity assessments, human oversight, accuracy documentation, and registration in an EU database before market placement.

• Risk classification — Unacceptable Risk (prohibited), High Risk (conformity assessment required), Limited/Minimal Risk
• Prohibited practices — social scoring, real-time biometric surveillance, subliminal manipulation, and emotion recognition at work
• High-risk requirements — risk management, data governance, technical documentation, human oversight, and accuracy
• GPAI model obligations — general-purpose AI models (GPT-class) face transparency and safety evaluation requirements
• Enforcement — national market surveillance authorities plus an EU AI Office for GPAI model oversight

NIST AI RMF 1.0: The Voluntary US AI Risk Management Standard

The NIST AI Risk Management Framework version 1.0, published in January 2023, provides a voluntary, flexible framework for managing AI risks — structured around four core functions: Govern, Map, Measure, and Manage. The Govern function establishes organisational policies, culture, and accountability for AI risk. Map identifies the context, risks, and potential impacts of specific AI systems. Measure quantifies AI risks through metrics, testing, and evaluation. Manage implements controls, monitors performance, and responds to AI-related incidents. The framework is designed to complement rather than replace existing AI regulations and is being widely adopted by US federal agencies and private sector organisations as a baseline for responsible AI deployment.

• Govern — organisational context, culture, accountability, and risk tolerance for AI development and deployment
• Map — identification of AI system context, purpose, risk categorisation, and impacted stakeholders
• Measure — AI risk quantification through testing, evaluation, red-teaming, and impact assessment
• Manage — risk treatment prioritisation, controls implementation, incident monitoring, and response
• AI RMF Playbook — supplementary guidance with suggested actions for implementing each framework function

MITRE ATLAS: Adversarial Threat Landscape for AI Systems

MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is an AI-focused analogue to MITRE ATT&CK — a curated knowledge base of real-world adversarial attacks against machine learning systems. ATLAS documents attack techniques observed in actual AI system compromises including reconnaissance techniques for model discovery, evasion attacks (adversarial examples that fool models), poisoning attacks (corrupting training data to manipulate model behaviour), extraction attacks (stealing model weights or functionality), and functional attacks (manipulating model outputs to achieve attacker goals). Security teams use ATLAS to evaluate their AI systems' resilience and design detection rules for AI-specific attack patterns.

• Reconnaissance techniques — methods attackers use to discover AI model APIs, training pipelines, and model specifications
• Evasion attacks — adversarial examples, white-box and black-box attacks that cause misclassification
• Poisoning attacks — training data corruption and backdoor attacks that manipulate model behaviour at inference
• Model extraction — techniques for stealing model functionality, weights, or training data through API queries
• Case studies — real-world AI attack incidents documented with technique attribution and detection recommendations

ISO 42001 and Organisational AI Management Systems

ISO/IEC 42001:2023, published in December 2023, is the first international standard specifying requirements for an Artificial Intelligence Management System (AIMS) — providing organisations with a framework for responsible AI development and deployment analogous to what ISO 27001 provides for information security. ISO 42001 covers AI policy, risk assessment specific to AI systems, impact assessment, operational controls for AI development pipelines, and performance evaluation. Organisations seeking to demonstrate third-party-verified AI governance can pursue ISO 42001 certification — and several EU AI Act high-risk compliance pathways are expected to reference ISO 42001 as a harmonised standard. Certification is available through accredited certification bodies.

• AI policy and leadership — executive commitment, AI governance policy, and organisational roles for AI oversight
• AI risk assessment — structured assessment methodology for identifying and evaluating AI-specific risks
• AI impact assessment — assessment of societal, ethical, and safety impacts of AI system deployment
• Operational controls — requirements for AI data governance, model development, testing, and deployment pipelines
• Certification pathway — third-party certification demonstrating conformity to ISO 42001 AIMS requirements