AI Security Documentation

This section provides a structured reference for securing AI systems across the lifecycle: from model development through deployment and ongoing operations. Every page cites final, in-force instruments — NIST publications, ISO standards, EU regulations, OWASP and MITRE frameworks — with resolvable URLs and specific provisions.

The landscape moves quickly. This documentation is updated as standards evolve; the revision history for each page is available via Git.

Standards & Frameworks

The normative foundations for AI security and governance.

• NIST AI Risk Management Framework (AI RMF 1.0) — Core, Profiles, and Playbook structure
• NIST Generative AI Profile (AI 600-1) — GAI-specific risk mapping and controls
• ISO/IEC 42001:2023 — AI management system requirements and Annex A controls
• EU AI Act — Regulation (EU) 2024/1689: prohibited practices, high-risk obligations, GPAI rules, and phased enforcement
• OWASP LLM Top 10 (2025) — LLM01–LLM10 taxonomy with mitigation patterns
• MITRE ATLAS — Adversarial Threat Landscape for AI Systems: tactics, techniques, and case studies
• Standards Comparison — Mapping overlaps and gaps across AI RMF, ISO 42001, EU AI Act, and OWASP

Threat Taxonomy

Catalogue of attack vectors organised by MITRE ATLAS alignment where applicable.

• Prompt Injection — Direct, indirect, and stored injection; boundary confusion; data exfiltration paths
• Jailbreaks — Role-play, encoding, multi-turn, and many-shot techniques; defence evasion
• Adversarial Attacks — GCG, PAIR, TAP, and gradient-based methods against open and closed models
• Data Poisoning — Clean-label, backdoor, and availability poisoning; supply-chain vectors
• Model Extraction — Query-based extraction, distillation, and membership inference
• Supply-Chain Attacks — Compromised dependencies, model hubs, and CI/CD pipelines
• Agent Hijacking — Tool misuse, context window manipulation, and multi-agent delegation abuse

Defensive Architectures

Practical controls mapped to threat vectors and standards requirements.

• Input Guardrails — Prompt validation, instruction hierarchy, and classifier-based filtering
• Output Filtering — PII redaction, groundedness checks, and refusal enforcement
• RAG Security — Vector store access control, retrieval authorisation, and context injection defence
• Runtime Monitoring — Anomaly detection, drift alerts, and behavioural baselines
• Secure Deployment — TEEs, confidential computing, model signing, and supply-chain verification
• Defence in Depth — Layered control strategy aligning NIST AI RMF functions to technical controls

Governance & Operations

Operationalising standards into auditable processes.

• AI Risk Assessment — AI RMF Map function: context, stakeholder, and risk identification methodologies
• AI Incident Response — Detection, containment, eradication, and post-incident reporting per ISO 42001 Annex A
• Audit & Evidence — Artefact collection, model cards, data sheets, and conformity assessment preparation
• Transparency & Documentation — System cards, dataset documentation, and EU AI Act Art. 50/53 disclosure obligations

Engagement

• Request an AI Security Assessment — Scope, methodology, delivery via NestVault365, and contact