1. Purpose
This standard defines the minimum set of governance, technical, and organisational controls required to protect the organisation's information security when artificial intelligence systems are used, including machine learning models, LLM-based solutions, RAG architectures, agentic scenarios, and AI-enabled cybersecurity tools.
It is intended to function as an internal normative document. It defines what must be done, who is responsible, what evidence must exist, and which legal acts and standards must be used when planning, building, deploying, operating, changing, and retiring AI systems.
2. Scope
- Creation, training, fine-tuning, testing, evaluation, and deployment of AI systems.
- MLOps / LLMOps processes, including model registries, CI/CD, inference services, prompt templates, RAG, and agentic workflows.
- Cloud, HPC, Kubernetes, OpenStack, and other infrastructures where data, model weights, vector indices, logs, and experimental artefacts are stored.
- AI-based cybersecurity functions such as fraud detection, phishing detection, sensitive information leakage detection, anomaly detection, and threat detection solutions.
- Both internally developed AI systems and third-party AI systems acquired, licensed, rented, or consumed through APIs.
If AI systems process personal data, profile individuals, generate recommendations that may have significant impact on individuals, or are integrated into products or services placed on the EU market, additional requirements from GDPR, the AI Act, the Cyber Resilience Act, and—depending on the sector—NIS2 or DORA must also be considered.
Not all legal instruments apply equally to all organisations. For this reason, this standard distinguishes between:
- the general baseline regulatory layer,
- the AI-specific regulatory layer,
- the sectoral / conditional layer.
3. Terms and Abbreviations
| Term | Definition |
|---|---|
| LA | An internal client term for an AI / algorithmic / digital information environment. If the organisation uses another official expansion, only this terminology section should be changed. |
| CAISO | Chief Artificial Intelligence Security Officer — a horizontal AI security governance function responsible for oversight of AI security principles, controls, and decision accountability. |
| AI system | A software or software-hardware system that generates predictions, content, recommendations, decisions, or actions according to predefined or adaptive objectives. |
| GPAI | General-purpose AI model that can be integrated into a broad variety of horizontal or vertical solutions. |
| MLOps / LLMOps | Governance, security, versioning, testing, deployment, and monitoring practices for AI models and their operation. |
| RAG | Retrieval-Augmented Generation — an architecture in which a generative model relies on external data sources, documents, or vector databases. |
| DPIA | Data Protection Impact Assessment under GDPR. |
| AI ISA | AI System Impact Assessment — an assessment of the AI system's organisational, social, ethical, and safety impact. |