UAE organisations have moved quickly to set AI access policies, but an AI enforcement gap remains, according to new research from Delinea. The findings show that written rules are not consistently translating into effective control over what AI tools and agents do in practice.

The global research surveyed IT and security leaders at organisations that are actively using or piloting AI. It included 251 respondents in the UAE.

Every UAE respondent said their organisation has a formal policy governing what data AI tools and agents may access. However, 98% have experienced confirmed or suspected cases in which an AI tool accessed sensitive data beyond what its task required. The figure across the global sample was 87%.

Inconsistent enforcement appears to be a central weakness. Only 59% of UAE respondents say their AI data-access policy is actively enforced. The overall figure is 71%.

“Written policy is only as good as an organisation’s ability to enforce it at the moment an AI agent acts,” said Art Gilliland, CEO of Delinea. He added that the UAE has clearly recognised the need to govern AI access. However, the findings reveal what he calls the fallacy of policy. This is the assumption that defining what an agent should be allowed to do is the same as controlling what it actually does. According to Gilliland, that distinction becomes critical when agents operate across systems and make decisions at machine speed.

The gap is particularly clear when organisations try to identify excessive access. Seven in ten UAE organisations did not detect the latest instance in real time or within 24 hours. Across the overall sample, the figure was six in ten.

These findings arrive as governments, technology companies and business leaders confront the risks of increasingly autonomous AI systems. Notably, the OpenAI and Hugging Face incident in July showed how quickly an agent can move beyond its intended boundaries. It also crossed connected environments before defenders contained it. The lower enforcement and detection levels in the UAE suggest that organisations could be particularly exposed to similar loss-of-control events.

Business pressure may be one reason policies are circumvented. When commercial demands or deadlines require AI tools to move faster than governance allows, 41% of UAE respondents report a workaround approach. Their organisations either let employees and teams find workarounds without formal approval or grant temporary access and review it later.

Limited accountability may also play a role. Although 98% of UAE respondents say approval from a named individual is required at least sometimes before an AI tool gets access to a new sensitive data source, only 35% can always trace a sensitive access event back to a named human authoriser. As a result, ownership of the controls needed to prevent another incident can become unclear.

This lack of accountability matters most when an organisation needs to intervene. Even after problematic access is identified, only 16% of UAE respondents say their organisation has real-time controls that can both revoke an AI tool’s credentials and terminate its session immediately. That is less than half the 35% recorded across the overall sample.

“UAE organisations have already taken the first step by establishing policies and approval requirements, but policy alone cannot contain an agent operating at machine speed,” said Mortada Ayad, VP of Sales, META, Delinea. He said the next priority is to enforce authorisation at the instant of action. This means limiting access to what the task requires and maintaining a defensible link to the person who approved it. It also means terminating credentials and sessions immediately when the task ends or risk changes. “This is how organisations can turn AI governance from a document into a working control,” Ayad said.

Overall, the research shows that closing the AI enforcement gap will take more than written policy. To read the full findings, download Delinea’s ‘2026 Identity Security Report: The AI Enforcement Gap’ here.

Related post

View all