Secure AI deployment, enterprise discipline.

What makes an AI deployment secure?

The main attack is well documented: prompt injection is ranked first (LLM01) in the OWASP Top 10 for LLM Applications 2025. The UAE’s data law, Federal Decree-Law No. 45 of 2021, applies to personal data processed in the UAE whether the processor is local or abroad.

1. Local models for sensitive data. Open-weight models (such as Llama or Mistral) running on infrastructure you control. Slightly less capable than frontier models, with complete data control.
2. Zero-retention API terms. Where a public model is the right tool, enterprise agreements that exclude your data from retention and training, over encrypted connections with minimal payloads.
3. Client-side masking. Names, contact details and account numbers are replaced with tokens before any external call and restored locally afterwards.
4. Audit logs. Every request, response and automated decision is logged with time, model and masked inputs, kept as long as your obligations require.
5. Deletion on request. Conversations, training data and caches can be purged, as the UAE PDPL requires.

When should you use a local model?

• Legal and compliance: contracts, case research, privileged material.
• Finance: client portfolios, transactions and tax planning.
• Health-adjacent: insurance claims and benefits.
• Private clients: family office, relocation and concierge data.

When is a public AI API fine?

Use of AI is near universal now: nearly nine in ten organisations in McKinsey’s 2026 survey use it in at least one function, so the question is how, not whether.

“By default, we do not use data from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu…”

OpenAI, Business data privacy, security and compliance (the same default applies to the API platform)

• Drafting, research and coding without client data.
• Customer-facing chat using only what customers choose to share.
• Analysis of public information.

A hybrid example

For a Dubai luxury services business:

• Customer chat: public model, enterprise terms, masking.
• Internal knowledge base: local model.
• Proposal drafting: masked locally, written by a public model, unmasked locally.
• Private-client portfolio review: local model only, no external calls.

Implementation discipline

• Sandbox first, with synthetic or anonymised data until the pilot passes.
• Human review where errors are costly.
• Regular reviews of workflows, and penetration testing for sensitive systems.
• Documented data flows for every workflow: inputs, storage, retention, access and deletion.

Which standards apply?

• UAE Personal Data Protection Law (Federal Decree-Law No. 45 of 2021).
• GDPR-equivalent practice where EU residents’ data is involved. See the EU AI Act guide.
• ISO/IEC 27001 and SOC 2 controls where you need certification-ready practice.
• The OWASP Top 10 for LLM applications, covering prompt injection and data leakage.