Malaysia Bars Government AI Vendors From Training Models on Public Data

Government contracts bar AI training and require data deletion

Malaysia prohibits companies providing artificial intelligence services to the government from using government or public data to train their models, Digital Minister Gobind Singh Deo said in a written parliamentary reply on October 20, 2026. The restriction appears in all government AI procurement contracts, which also require providers to delete the data after processing.

Gobind was responding to Ahmad Fadhli Shaari, the Perikatan Nasional member of Parliament for Pasir Mas, who asked how information handled by cloud and AI companies, including foreign providers, remains under Malaysian control. His answer set out controls covering storage locations, encryption, access to government systems and the reuse of information by AI suppliers.

The practical distinction is between using information to complete a government task and using it to develop a supplier's AI model. An agency may need an AI system to process information for a particular service. Under the contractual restriction Gobind described, the provider cannot then use that information as material for training its models or retain it once processing is complete.

Training is the process through which an AI model learns patterns from examples. Processing is the use of a system to perform a task with information it receives. Keeping those activities separate gives agencies a contractual basis for controlling what happens to information after they send it to a provider.

Gobind also said the Digital Ministry is drafting an AI Governance Bill to address accountability throughout the life of an AI system. The measures already described operate through government contracts, security directives and public sector guidelines, while the proposed legislation would establish a broader framework for responsibility.

Where each category of government data can go

Malaysia's storage rules depend on how sensitive the information is. Gobind said material classified as secret or top secret must remain in government data centres. Confidential and restricted information may be stored only in private or hybrid cloud environments based in Malaysia. Only open, nonsensitive government data may be stored in cloud services overseas.

A cloud service stores or processes information on computing infrastructure accessed through a network. A private cloud is reserved for a particular organisation, while a hybrid cloud combines separate computing environments. Under the arrangements Gobind described, the classification of a government record determines which storage options an agency can use.

• Secret and top secret information: government data centres, rather than cloud services.
• Confidential and restricted information: private or hybrid cloud environments located in Malaysia.
• Open, nonsensitive information: eligible for storage in cloud environments overseas.

For an agency selecting a provider, the first decision is therefore about the information itself. A supplier's ability to process data does not settle where that data may be hosted. The agency must match its storage arrangements to the security classification before using the service.

Encryption keys remain under government control

Government data must be encrypted both while stored and while moving between systems, in accordance with the National Cryptography Policy, Gobind said. These protections are commonly described as encryption at rest and encryption in transit.

Encryption turns readable information into coded content. A cryptographic key provides the means to recover the readable version. For classified systems and data, Gobind said the government owns and controls those keys, with the aim of preventing a foreign service provider from independently reading the contents, including in response to a request from another country's authorities.

The storage rules and the key controls address different parts of the same problem. Storage rules determine where information can reside. Key ownership determines who holds the means to read encrypted content. For classified information, the government says it retains that control alongside the restrictions on hosting locations.

Who can enter government systems?

Companies appointed to provide government services must undergo security screening. Gobind said access to government systems is assigned according to a person's role, requires authentication using more than one factor and is recorded for audits.

Access based on roles means that a person's permissions depend on the work they are authorised to do. Authentication using multiple factors requires more than a password alone, such as an additional security check. Audit logs create a record of access that can be examined later.

These requirements extend protection beyond the location of a server. A system hosted in Malaysia still needs controls over the people who can enter it, the information they can reach and the record of their activity. Vendor screening, restricted permissions and access logs form part of the government's conditions for handling its data.

Agencies must reduce data before AI processing

The Public Sector AI Adoption Guidelines, known as GPAISA, require agencies to minimise the amount of data used and conceal or anonymise information before it is processed by AI systems, Gobind said. Data minimisation means using only what is needed for the task. Concealing or anonymising data reduces the exposure of information that could identify a person.

The National Digital Department's public sector AI guidelines cover compliance with Malaysian laws and policies, seven principles of AI ethics, the responsibilities of people and organisations involved in AI, adoption procedures and risk management. Their appendices include public sector use cases from Malaysia and abroad, together with a self assessment for ethics compliance.

The department's AI initiative report describes GPAISA as a reference for public sector agencies to ensure that AI is used ethically, transparently and responsibly. It says the guidelines focus on improving service delivery, strengthening government operations and encouraging innovation based on new technologies. During 2025, the department provided 40 technical advisory services to government agencies to help coordinate AI policies and practices.

The framework assigns responsibilities to agencies as well as suppliers. Before information reaches an AI service, an agency must consider how much data the task needs and how to protect identifying details. The contract then governs the supplier's use and deletion of the information it receives.

The ethics assessment and risk management sections also place data handling within the wider decision about whether and how to use an AI system. Procurement is one stage of that decision; preparation of information and responsibility for the resulting service remain part of the agency's work.

A framework launched in February 2025

The Digital Ministry's guidelines launch statement records that GPAISA was launched on February 27, 2025, at Hotel Le Meridien in Putrajaya. The National Digital Department developed the guidelines with the Malaysia Digital Economy Corporation as a dedicated reference for public agencies.

Gobind officiated at the launch alongside Deputy Director General of Public Service for Operations Dato' Mohd Shaiful bin Ibrahim. The event brought together government department heads, chief digital officers, public sector technology managers and industry participants, with sessions on government AI use cases, ethics and governance.

The guidelines contain six chapters and three appendices. Their publication preceded the October 2026 parliamentary explanation of contractual restrictions and the drafting of the proposed AI law. The main dates are:

• 2024: Amendments to the Personal Data Protection Act 2010, cited by Gobind in explaining transfers of personal data overseas.
• February 27, 2025: Launch of the public sector AI guidelines in Putrajaya.
• October 20, 2026: Gobind's written parliamentary reply explains the government's AI contract conditions, storage rules and work on an AI Governance Bill.

Private sector transfers follow a separate legal route

For personal data handled by private businesses, Gobind pointed to Section 129 of the Personal Data Protection Act 2010, as amended in 2024. He described the requirements for transfers overseas as including consent from the person concerned and laws in the receiving jurisdiction comparable to Malaysia's PDPA.

A transfer across borders can occur when a business sends personal information to a cloud service located abroad. Gobind's explanation places those transfers under the personal data protection framework, while government information is subject to the storage classifications and security controls he outlined separately.

The difference matters for people whose information is handled by both public agencies and businesses. The government's procurement conditions govern suppliers working on its AI contracts. Private businesses handling personal information must consider the PDPA rules for transferring that information to another country.

What the proposed AI law would cover

The Digital Ministry is drafting an AI Governance Bill intended to hold technology providers accountable throughout an AI system's life cycle, Gobind said. He identified data quality, transparency and explainability among the areas it would address. No date was given for introducing the bill in Parliament, and the reply did not set out proposed penalties or a detailed enforcement structure.

Data quality concerns the information used by a system, including whether it is accurate and suitable for the task. Transparency concerns visibility into how a system is developed and used. Explainability concerns making its outputs or decisions understandable to people. Those subjects extend accountability beyond the initial purchase of an AI service.

The controls come as Malaysia expands its computing infrastructure. Plantation land redevelopment for data centres and solar parks has drawn major palm oil groups into the sector. Earlier figures put potential data centre electricity demand at at least five gigawatts by 2035, close to 20 percent of the country's generating capacity at the time, alongside a government target of 81 operating data centres by 2035.

For agencies and suppliers, that expansion makes the rules on permitted uses concrete. Building capacity to store and process information creates more options for digital services. Government contracts, data classifications, encryption keys and access controls determine which of those options agencies can use and what providers may do with the information.

Key Points

• Government AI contracts prohibit providers from using government or public data to train models and require deletion after processing.
• Secret and top secret information must remain in government data centres. Confidential and restricted information may use domestic private or hybrid clouds.
• Only open, nonsensitive government data may be stored in cloud environments overseas.
• The government controls encryption keys for classified systems and data, while vendors face security screening and access controls.
• Public agencies must minimise data and conceal or anonymise it before AI processing under GPAISA.
• An AI Governance Bill is being drafted to address accountability, data quality, transparency and explainability.