By Umair Bin Ahmed, Founder & Managing Director, Asteriqx Consulting Limited
Ask where a country's most important AI systems actually run, and the honest answer is often: somewhere else. The models are trained abroad, hosted on foreign cloud platforms, assembled from open-source components, and updated on schedules no local authority sets.
That is the quiet paradox of AI governance today. Organizations and governments are writing their own Responsible AI rules, yet much of the AI they want to govern sits in a supply chain they do not control. It is also where the newest cyber security risks are getting in.
Generative AI Rewrites the Risk Register
To a large language model, instruction and a piece of content are the same thing: text. A hidden line in an email or a supplier's document can be read as a command, which is why prompt injection tops the OWASP Top 10 for LLM Applications. AI agents raise the stakes further. An assistant that answers can leak data; an agent that can send, pay, or delete can do real damage once manipulated.
Controls have not kept pace. IBM's 2025 Cost of a Data Breach Report found that 13% of organizations had already suffered an attack on their AI models or applications. Of those, 97% lacked proper AI access controls. And 63% of organizations had no AI governance policy at all.
Most of the Risk Arrives Through Third Parties
Verizon's 2025 Data Breach Investigations Report found that third-party involvement in breaches had doubled to 30%.
A typical AI service depends on a foundation model provider, cloud host, vector databases, plug-ins, open-source libraries, and training data of uncertain origin. Each is a supplier. Few vendor questionnaires even ask which model a supplier uses, let alone where it is hosted.
In February 2024, researchers at JFrog found around 100 malicious models on Hugging Face, some able to plant a backdoor on the machine of anyone who loaded them. A poisoned model looks like any other file until it runs.
Concentration is the other danger. In July 2024, a faulty update from a single security vendor, CrowdStrike, crashed about 8.5 million Windows devices and disrupted airlines, hospitals, and banks worldwide. AI is consolidating around even fewer providers. One flawed model update could reach thousands of organizations at once.
Where Governance and Security Fail to Meet
Inside most organizations, AI governance and cyber security still run as separate conversations. Governance committees write principles on fairness, transparency, and accountability. Security teams write controls for access, monitoring, and incident response. The AI supply chain falls between them.
Governance may approve a vendor's AI tool as ethically sound without asking what data it can reach or what happens when its model changes. Security may assess the vendor's infrastructure without knowing that an AI model now sits inside the product. The result is accountability without control: the deployer answers regulators, while the vendor holds the levers.
Why This Is a Sovereignty Question
The same gap exists at national level. When a provider updates its model, behavior can change for every user in every country overnight. When export controls tighten, access to advanced chips and models can be restricted by decisions taken in another capital. None of these choices is made in Kuala Lumpur, Jakarta, or Manila.
Models trained mainly on data from elsewhere may also perform poorly in local languages and contexts. A country that cannot test the AI it uses cannot know whether that AI is safe, secure, or fair to its own citizens.
The region is responding. ASEAN's Guide on AI Governance and Ethics, expanded in 2025 to cover generative AI, calls on upstream developers to share important information with downstream deployers, and proposes common approaches to testing and incident reporting. Singapore's IMDA published a Model AI Governance Framework for Agentic AI in January 2026. In July 2026, Malaysia's National AI Office opened consultation on an AI Governance Bill that would reach systems used by Malaysian deployers, wherever they are hosted.
Yet a law only reaches as far as an organisation's ability to see into its suppliers.
Sovereignty Means Choose, Verify, Exit
The answer is not digital nationalism. Few countries can build every layer of the AI stack, and none should try.
A more useful definition is this: sovereignty is the ability to choose, verify and exit. Choose suppliers on informed terms, verify how their systems behave, and move away when trust breaks down. That is far stronger than Responsible AI built on principles alone.
Five Practical Moves for GRC Leaders
First, create one AI risk register, owned jointly by governance and security. Every AI system, built or bought, should have a business owner and a security counterpart, and be assessed for ethical and cyber risk together.
Second, map the AI supply chain, including fourth parties. Ask vendors for an AI bill of materials: the models, datasets, libraries and cloud regions behind their service, and their own key suppliers.
Third, rewrite contracts for AI. Require advance notice of material model changes, clear data residency and use terms, AI incident notification, and the right to audit or obtain independent assurance.
Fourth, test like an adversary, and test locally. Red-team supplied AI for prompt injection, data leakage, and misuse, in local languages, before deployment and after every significant update.
Fifth, measure concentration and plan the exit. Know how many critical processes depend on one model provider or cloud region. Keep data, prompts and evaluation sets portable, and rehearse switching for critical services.
Governments Hold the Biggest Lever
Public procurement is the strongest tool regional governments have. When agencies demand transparency, security testing and exit rights from AI suppliers, the market adjusts for everyone. Mutual recognition of AI assurance across ASEAN would give the region bargaining power that no single member has alone.
The Question Worth Asking
Responsible AI cannot be delivered by policies written in one place and executed by systems controlled in another. It depends on who we depend on, and on what we can see, test and change.
Boards and regulators should stop asking only, “Is our AI ethical?” and start asking a harder question: if our most important AI supplier were compromised, changed course, or walked away tomorrow, would we still be in control?
Sources (for editorial factchecking; not for publication)
IBM, 2025 Cost of a Data Breach Report
OWASP Top 10 for LLM Applications 2025 (Invicti summary)
Verizon, 2025 Data Breach Investigations Report
BleepingComputer, Malicious AI models on Hugging Face backdoor users’ machines
Microsoft, Helping our customers through the CrowdStrike outage
Rajah & Tann, Expanded ASEAN Guide on AI Governance and Ethics
Computer Weekly, Singapore debuts world’s first governance framework for agentic AI
Baker McKenzie, Malaysia: Public Consultation on the AI Governance Bill
Latest News
