Abu Dhabi is introducing new security protections for more than 50 government entities as authorities seek to prevent artificial intelligence systems from being manipulated into accessing sensitive information or carrying out unauthorised financial transactions.
The Department of Government Enablement (DGE) is developing layers of oversight for AI models and autonomous agents as government departments move towards AI-native operations. The measures include monitoring, auditing, restrictions on system activities and the ability to shut down models if they behave unexpectedly.
Najyb Al Maskari, executive director of the Government Cybersecurity Operations Sector at DGE, said the main concern was not simply that an AI system could make an error, but that attackers might manipulate it into performing actions that appear legitimate.
“Malicious attackers will tell the agent to do certain things, and the agent will carry out those tasks as if they were authorised,” Al Maskari said.
He warned that a compromised system could potentially be instructed to disclose financial records, access personal information or transfer money.
“I can tell you that we are securing all 50 entities,” he said, adding that the department oversees more than 50 government entities and is working towards establishing an AI-native government across them.
Al Maskari said the security framework must follow a principle of defence in depth, with multiple safeguards rather than reliance on a single protective measure.
These safeguards include monitoring models after deployment, reviewing their activities, applying restrictions through built-in guardrails and retaining the ability to shut down systems when necessary.
He said problems could emerge when AI models are developed and autonomous agents are built on top of them to perform tasks without sufficient supervision.
The warning comes as AI agents increasingly move beyond answering questions to accessing computer systems, executing instructions and completing complex workflows on behalf of users.
Dr Najwa Aaraj, chief executive of the Technology Innovation Institute, raised similar concerns at the AI Everything event last week. She warned that an AI agent operating inside a computer system without real-time monitoring could become a serious security threat.
“If you cannot, in real time, monitor the execution of this agent, this is a malware that got access to your machine,” Aaraj said.
She called for stronger standards governing AI agents, including clear identities for non-human systems and restrictions on the information and functions they can access.
Vladislav Tushkanov, head of the Kaspersky AI Research Center, said AI agents were increasingly attractive targets for cybercriminals because they may operate on devices connected to corporate networks and have access to credentials and confidential information.
He highlighted malicious AI “skills”, which provide agents with packaged instructions and tools, as a potential route for attacks. Criminals could hide harmful code in these packages or direct agents to download and execute malicious files.
Tushkanov described a developing technique known as “living off local agents”, in which attackers exploit AI tools already installed on a computer to search for valuable information.
He also warned that AI was enabling less sophisticated criminals to launch more attacks against more targets at greater speed.
For organisations adopting AI, he said conventional cybersecurity measures remained essential but needed to account for attacks involving automated agents. He recommended combining endpoint protection, threat detection and response capabilities, and threat intelligence.
Al Maskari said Abu Dhabi’s objective was to establish a secure, trusted AI-native government that places people at the centre of its operations.
“We are working on an AI-native government that is trusted by the citizens and residents of the UAE and is secure at all levels,” he said.
