UAE AI Regulation 2026: Security Testing Requirements for AI Systems

Update (October 2026): An earlier version of this guide was framed around a “UAE AI Act 2026” and its risk tiers. We could not verify that any such Act, or a September 2026 self-assessment deadline, exists in a primary source, so we have rewritten the guide around instruments we can verify. We also removed a PDPL deadline we could not confirm. The URL is unchanged.

UAE AI regulation in 2026 is not one law. It is a set of instruments that together expect AI systems to be governed, overseen and secure: the CBUAE AI guidance note for financial institutions, the new Federal Authority for AI and Data, the Dubai AI Seal for government suppliers, DESC’s security policy and testing rules, DIFC Regulation 10 and the PDPL. None names AI penetration testing, but all of them need testing to evidence.

If you searched for “UAE AI Act 2026”, you probably found confident articles describing a four-tier statute with fixed fines and a self-assessment deadline. We looked for the decree behind them and could not find one. This guide sticks to what we can source, and then gets practical about AI security testing.

What AI rules actually apply in the UAE in 2026?

Sources include the CBUAE press release, Morgan Lewis on the federal authority, Baker McKenzie on the Dubai AI Seal, Dubai Media Office and Mayer Brown on DIFC Regulation 10.

The honest summary: no single AI statute, but if your AI touches customers, personal data or government services, at least one of these already applies to you.

What does the CBUAE guidance mean for banks and fintechs?

The CBUAE note is the most concrete AI-specific instrument in the UAE so far. It applies to all licensed financial institutions and sets principles for governance and accountability, fairness, transparency and explainability, human oversight, and data management and privacy.

Read with a security lens, each principle needs evidence. Human oversight means proving the AI cannot take high-impact actions without approval, which you test. Data protection means proving the chatbot cannot be talked into leaking another customer’s data, which you test. Our CBUAE penetration testing guide covers the wider testing expectations for banks and payment firms.

What about Dubai government and the AI Seal?

Two gates apply to anyone selling AI into Dubai government. The Dubai AI Seal decides which AI companies entities should buy from. DESC sets the security rules, including the Dubai AI Security Policy and the Cyber Force scheme for who may deliver formal penetration tests (see our DESC guide).

With Dubai’s Agentic AI for Government Services Accelerator launched in October 2026, this has become urgent for vendors. Our DESC and AI Seal vendor checklist walks through it.

What is happening in Saudi Arabia?

If you serve KSA clients, the NCA AI Cybersecurity Guidelines are the document to watch. The NCA ran a public consultation from 5 July to 5 August 2026. The draft is organised around governance, defence, resilience and third-party security, and explicitly covers generative and agentic AI. We could not confirm a final version or whether it will be binding, so plan for it without treating it as law yet.

How much testing does each AI system need?

Without a statutory tier list, use a simple risk-based lens. It mirrors how supervisors think and how the EU AI Act works for firms with European exposure.

The point of the lens is budget. You do not red-team an internal summariser. You absolutely red-team the agent that can move money.

For high-risk systems, the test plan should cover the OWASP LLM Top 10 and the OWASP Top 10 for Agentic Applications: direct and indirect prompt injection, tool misuse, identity and privilege abuse, memory poisoning and data leakage. If agents connect through MCP servers, include those too; see our MCP server security assessment. The AI agent field guide explains the mechanics.

What evidence do UAE supervisors and auditors expect?

Across CBUAE, DFSA, VARA, DESC and PDPL-driven reviews, what closes questions is a repeatable evidence package, not a vendor attestation:

1. Scope and methodology: which systems, which standards, which attack classes.
2. Findings with severity, each reproducible.
3. Proof of concept: shown impact, not theoretical risk.
4. Remediation record: what was fixed, by whom, when.
5. Retest confirmation, dated.

That package also maps to most AI governance frameworks, so one programme can serve several regulators.

A practical way to start

1. Inventory your AI: every system using an LLM or automated decisions, with the data and actions it touches.
2. Map instruments: CBUAE if you are licensed, DESC and AI Seal if you sell to Dubai government, DIFC Regulation 10 if you are in the DIFC, PDPL for personal data, NCA if you serve KSA.
3. Rank by risk using the lens above.
4. Test proportionally and package the evidence.
5. Re-test on change: new models, new tools and new data sources reopen risk.

Budgeting help is in our AI red teaming cost guide.

Getting your AI tested

We will keep this page updated as the Federal Authority for AI and Data publishes standards and as the NCA finalises its guidelines. The requirement underneath will not change: if your AI touches customers, personal data or government services, you need to prove it is secure.

pentest.ae offers a fixed-scope LLM penetration test for a single application, a broader AI security assessment, and a full agentic red team exercise for autonomous systems, each with a retest and an evidence package built for UAE supervisors.