Skip to main content
ISO 27090

ISO 27090 AI Security Guide: Threats, Controls & EU AI Act

A practical practitioner guide to ISO/IEC 27090 AI security threats, prompt injection, data poisoning, and bridging EU AI Act Article 15 compliance.

Overview: The Arrival of ISO/IEC 27090:2026

The rapid enterprise adoption of Large Language Models (LLMs), agentic frameworks, and generative AI systems has introduced an attack surface that traditional cybersecurity frameworks (such as ISO/IEC 27001 Annex A or NIST CSF) were never designed to address.

Vulnerabilities like prompt injection, training data poisoning, model inversion, and adversarial evasion bypass perimeter firewalls and endpoint protection.

Formally published as ISO/IEC 27090:2026 (Cybersecurity — Artificial intelligence — Addressing security threats and failures in artificial intelligence systems), this international guidance provides the first standardized global framework for classifying AI-specific threats and deploying technical countermeasures across the AI lifecycle.

Crucially for European organizations, ISO/IEC 27090 provides the exact technical and operational baseline needed to satisfy Article 15 (Accuracy, Robustness and Cybersecurity) of the EU AI Act.


Why ISO/IEC 27090 Matters for EU AI Act Compliance

Under Article 15 of Regulation (EU) 2024/1689 (EU AI Act), high-risk AI systems must be designed and developed with appropriate levels of accuracy, robustness, and cybersecurity:

“High-risk AI systems shall be resilient against attempts by unauthorized third parties to alter their use, outputs or performance by exploiting system vulnerabilities.” — EU AI Act, Art. 15(3)

Article 15 specifically calls out threats such as data poisoning, adversarial examples, model flaws, and jailbreaking. While the EU AI Act specifies what legal obligations must be met, it does not detail how engineers and security teams should configure technical defenses.

ISO/IEC 27090 fills this exact gap:

EU AI Act Article 15 RequirementISO/IEC 27090:2026 Countermeasure DomainOperational Control Document
Art. 15(3)(a) Data Poisoning ResiliencyTraining Data Provenance & Anomaly Detection (Cl. 6.2)Data Pipeline Integrity Verification Register
Art. 15(3)(b) Adversarial Attacks & EvasionModel Robustness & Perturbation Testing (Cl. 7.4)AI Red-Teaming & Stress-Testing Protocol
Art. 15(3)(c) Prompt Injection & JailbreakingInput Sanitization, Guardrails & LLM Filtering (Cl. 8.1)Generative AI Prompt Gateway Architecture
Art. 15(4) Confidentiality & Model TheftModel Extraction & Membership Inference Defense (Cl. 7.6)Differential Privacy & Query Rate Throttling Standard

The ISO/IEC 27090 AI Threat Taxonomy

ISO/IEC 27090 breaks AI security down into four distinct vulnerability and attack tiers:

1. Data-Tier Threats

  • Data Poisoning: Malicious tampering with pre-training, fine-tuning, or retrieval-augmented generation (RAG) data to bias outputs or embed backdoors.
  • Training Data Leakage: Memorization vulnerabilities where LLMs regurgitate sensitive PII or trade secrets through targeted extraction prompts.

2. Model-Tier Threats

  • Adversarial Evasion: Crafting subtle input perturbations that trigger catastrophic misclassification while appearing normal to human auditors.
  • Model Inversion & Reconstruction: Reverse-engineering model weights or reconstructive training data through repetitive inference queries.
  • Model Theft: Copying proprietary models via high-volume synthetic API query cloning.

3. Application & Prompt-Tier Threats

  • Direct & Indirect Prompt Injection: Overriding system instructions via untrusted user inputs or poisoned third-party data retrieved by search agents.
  • Jailbreaking: Bypassing ethical alignment guardrails to generate harmful or unauthorized payloads.
  • Insecure Output Handling: Unsanitized LLM responses fed directly into downstream SQL databases, APIs, or shell scripts.

4. Infrastructure & Supply Chain Threats

  • Compromised Base Models: Integrating unverified open-source model checkpoints containing serialized malicious code (e.g., malicious pickle or safetensors payloads).
  • RAG Vector Database Tampering: Inserting unauthorized embeddings into semantic knowledge stores.

How to Implement ISO/IEC 27090 in 5 Steps

Step 1: Establish AI Asset & Threat Scoping (Cl. 4-5)

Document every foundation model, internal model, fine-tuning dataset, RAG repository, and external AI API in an AI System Asset Register. Determine exposure tiers (internal, partner-facing, or public-facing).

Step 2: Implement Defense-in-Depth for LLM Inputs & Outputs (Cl. 8)

Deploy multi-layered prompt filtering gateways. Separate trusted system prompts from untrusted user inputs using clear delimiter architectures. Validate and sanitize model outputs before execution.

Step 3: Enforce Continuous AI Red-Teaming (Cl. 7)

Conduct regular adversarial red-teaming exercises specifically targeting model jailbreaking, prompt exfiltration, and supply chain integrity. Log all testing results in an auditable register.

Step 4: Map to ISO/IEC 42001 and ISO/IEC 27001

If your organization operates an ISO 27001 ISMS or is pursuing ISO 42001 (AIMS) certification, link ISO 27090 controls directly to Annex A controls (such as A.8.25 secure development lifecycle and A.8.28 secure coding).

Step 5: Prepare Audit-Ready Documentation

Regulators and enterprise procurement teams will request documented proof of AI security. Ensure your policy statements, threat registers, and incident runbooks are codified in formal DOCX and PDF deliverables.


Accelerate Your Implementation

Don’t spend months drafting policies from scratch. The ISO/IEC 27090 AI Security Threats & Mitigation Blueprint provides:

  • 38-page editable Word document (.docx)
  • Complete AI Threat Taxonomy & Vulnerability Register (Excel)
  • Full crosswalk to EU AI Act Article 15 and ISO/IEC 42001
  • Incident response runbook for AI failure modes
  • Instant download with 7-day money-back guarantee

Explore the ISO/IEC 27090 Blueprint or get the Complete AI Security & EU AI Act Suite.

Related Blueprint

ISO/IEC 27090 AI Security Threats & Mitigation Blueprint

The definitive implementation blueprint for ISO/IEC 27090:2026. Identifies and mitigates AI-specific threats across data…

View Blueprint — €89