Vietnam Proposes Five Data Risk Levels, Stricter Overseas Transfers and AI Safeguards

Five risk levels would determine how Vietnam protects data

Vietnam's government presented a Data Security Law to the National Assembly on October 26, 2026, proposing five risk levels, stricter checks on sensitive data sent overseas and safeguards for information used by artificial intelligence. The draft contains six chapters and 29 articles, with just two administrative approval procedures, both focused on data at the highest risk levels.

Minister of Public Security General Luong Tam Quang, acting under the Prime Minister's authorization, presented the proposal during the second session of the 16th National Assembly. Businesses and other data controllers would initially assess their own data, apply protections matching its risk and remain accountable for that assessment. Authorities would mainly check compliance afterward.

The government proposes July 1, 2027, as the law's effective date. Deputies were scheduled to discuss it in groups on the afternoon of October 26. The legislation remains a proposal, with parliamentary reviewers seeking tighter controls on the most sensitive information, clearer assessment procedures and stronger protection for biometric and genetic data.

The bill would reach beyond organizations based in Vietnam. It covers Vietnamese agencies, organizations and individuals; foreign organizations and individuals in Vietnam; and those abroad processing data under Vietnamese jurisdiction or acting against Vietnam's data security.

For businesses, the central change would be a graduated system: routine processing would face basic safeguards and little paperwork, while activities capable of causing serious harm would require assessments, audits or approval before they begin. For individuals, the proposal adds security duties around information as it moves through systems, suppliers and national borders.

Why the government wants a separate law

Luong Tam Quang told lawmakers that Vietnamese information systems recorded about 552,000 cyberattacks in 2025, with 5,230 agencies and businesses suffering harm. He also cited roughly 300 detected cases of data being advertised for sale, involving more than 1.7 million records, and a theft involving more than 160 million customer records at a credit institution.

The minister said 60% to 70% of data leaks originated with personnel inside organizations. That makes employee and contractor access a central issue alongside protection against outside attackers. Restricting who can see information, monitoring access and removing permissions when necessary are practical parts of the proposed framework.

The government argues that existing rules address different parts of the problem. The Data Law governs activities such as creating, managing, sharing and classifying data. Personal data protection legislation governs processing conditions and individuals' rights. Cybersecurity legislation protects information systems against threats.

The proposed Data Security Law would focus on the information itself throughout collection, storage, processing, transmission and destruction. Its policy foundations include Politburo Resolution No. 57-NQ/TW of December 22, 2024, on science, technology, innovation and national digital transformation.

What the five levels would mean

Article 8 would rank data according to the likelihood of an incident and the damage it could cause to national defence, national security, public safety and the lawful rights of organizations and individuals. Level 1 is the lowest and level 5 the highest.

The proposed categories give businesses a starting point for identifying their obligations:

• Level 1 includes public and open data, with low incident likelihood and consequences.
• Level 2 includes internal information, business data and personal data outside the higher category, with medium risk.
• Level 3 includes important data and sensitive personal data processed at a large scale, subject to thresholds set by the government.
• Level 4 generally covers core data associated with very high risk and potentially severe harm.
• Level 5 includes core defence and security data and information used to operate or control particularly important facilities linked to national security.

The risk rating would operate alongside classification under the Data Law. Organizations and individuals would retain their lawful ownership, management and use rights over their information. The security level would determine the protection required.

Where paperwork and advance checks would begin

The government's presentation identifies two administrative procedures: appraisal of design documents for systems at levels 4 and 5 under Article 11, and appraisal and approval of overseas transfers at those levels under Article 19.

Level 1 would involve no administrative procedure. Level 2 controllers would prepare a shortened protection plan using a template and keep it themselves. Level 3 controllers would assess their own activities without submitting them for appraisal or notification. From level 3 upward, impact assessments and periodic security audits would be required.

Systems processing level 4 or 5 data would need data security assessment before entering official operation, coordinated with cybersecurity appraisal. An impact assessment examines what could go wrong and who could be harmed; an audit checks whether the promised protections actually work.

Household businesses, small and micro enterprises, and innovative startups handling routine data would generally follow basic protection guidance. Their obligations would still depend on the risk of their processing, making the type and scale of information they handle a practical compliance question.

The draft also seeks to reduce duplication. Where several laws require checks on the same subject and risk, an organization would use one set of documents and a single point of contact. Valid results obtained under other laws would be recognized. AI data security assessment would be folded into the risk assessment process under AI legislation.

Overseas transfers would face three different routes

Data sent abroad would receive different treatment depending on its level. Levels 1 and 2 would require no transfer procedure under this law. Level 3 controllers would assess the transfer themselves and retain records. Levels 4 and 5 would require review and approval from a competent authority before transfer.

Cloud services, international platforms and companies operating across several countries would therefore need to identify which information moves overseas and its assigned risk. Existing obligations under other relevant laws would continue to apply.

The National Assembly's Committee on National Defence, Security and Foreign Affairs broadly supported controls based on risk without a general requirement to keep all data inside Vietnam. It proposed a stricter rule for level 5: overseas transfers would be prohibited unless the Prime Minister decided otherwise.

The committee also proposed restricting overseas transfers of decryption keys and cryptographic parameters for systems at levels 4 and 5. A decryption key unlocks encrypted information, so control over that key is a separate security concern from the location of the stored data.

Other proposals would clarify who can restrict or temporarily suspend services operating across borders and require notification when control changes at a business managing data at level 4 or higher. The draft also addresses direct foreign requests for information through mechanisms intended to require lawful channels, including judicial assistance arrangements.

AI safeguards would start with training data

Organizations researching, developing or deploying AI that processes data posing high risks would have to build security into the design stage. They would need to establish the lawful origin and integrity of training data and information derived from it.

The proposal addresses data poisoning, unauthorized extraction and human control over automated decisions. Data poisoning means deliberately corrupting the information used to train or operate an AI system, potentially causing unreliable or manipulated results. Unauthorized extraction involves obtaining protected information without permission.

Human intervention and oversight would have to remain possible for automated decisions under the applicable law. That requirement connects protection of the underlying information with the decisions a system makes using it.

Authorities would be barred from demanding AI source code or model weights except for criminal investigations. Source code contains a program's instructions; model weights are the numerical settings an AI system learns during training. Both can contain valuable commercial knowledge.

AI businesses would use the assessment process under AI legislation rather than prepare a separate data security file or undergo a separate appraisal. This joins the proposed security checks to an existing regulatory process while preserving specific duties over the data itself.

Suppliers would become part of the security plan

Data controllers would have responsibilities extending across suppliers, intermediaries and other parties involved in processing information. Before connecting a product or service, they would assess the provider's risks and limit its access to what is needed.

Contracts would need to specify confidentiality duties and consequences for data leaks. Controllers would also monitor access during the relationship and revoke it when risks or incidents are detected. Remote supplier access to systems handling data at higher risk would face tighter controls.

The scope extends to critical digital infrastructure, national data infrastructure, large scale processing and systems connecting digital controls with physical equipment. In those settings, damaged or manipulated data can affect the operation of machinery and facilities as well as expose information.

The committee asked for responsibilities to follow each party's actual decision making power and control over data. That would help distinguish the duties of a controller choosing how information is used from those of a provider carrying out a narrower service.

Emergency powers would have legal limits

The draft would establish monitoring, early warnings, incident reporting and coordinated responses. Revisions described in the government's presentation also include informing affected people when incidents occur and limiting the scope of temporary suspensions of data transfers.

Emergency technical measures would require legal grounds, an authorized decision maker, a defined scope and a specified duration. Authorities would be directed to prefer the measure that restricts rights least and end it when its grounds cease to apply.

Affected organizations and individuals would retain rights to complain, bring legal proceedings and seek compensation under the law. Parliamentary reviewers also sought clearer conditions and compensation arrangements when authorities mobilize privately held personnel, equipment or other resources.

The draft encourages people responsible for data security and independent specialists to detect, test and report vulnerabilities in good faith, with protections for their identities and lawful interests and corresponding liability exemptions. The committee additionally proposed incident response exercises for systems at levels 4 and 5.

Reviewers want tighter rules for the highest risks

Senior Lieutenant General Le Tan Toi, chairman of the National Assembly's Committee on National Defence, Security and Foreign Affairs, presented the appraisal report. The committee supported the need for the law and considered the proposal ready for parliamentary consideration.

Reviewers nevertheless sought clearer criteria for each risk level. Some proposed that specialized authorities regularly review and update designations, with a stricter process for level 5 rather than leaving that decision to controllers themselves.

The committee also sought a clearer rule that the risk of the data determines the protection required of the processing system. It requested stronger safeguards for biometric and genetic information, such as information about bodily characteristics or inherited traits.

Procedural questions remain under discussion, including appraisal deadlines for systems at levels 4 and 5 that fall outside cybersecurity appraisal. The committee requested clearer boundaries between security cryptography, government cipher systems and civilian cryptography, along with incident coordination rules for military and government cipher systems.

The Ministry of Public Security would lead state management of data security, the Ministry of National Defence would oversee military data and information systems, and the Government Cipher Committee would manage cipher information systems. No new agency would be created.

What changed and what happens next

The proposal presented to lawmakers expanded from 25 to 29 articles, removed a planned National Data Security Governance Centre and dropped fines calculated as a percentage of revenue. An earlier description dated September 7 had outlined four data levels, possible fines of up to 5% of the previous fiscal year's revenue and domestic storage requirements for critical and core data. The parliamentary proposal instead uses five risk levels and differentiated transfer controls.

The government also removed a fixed deadline for switching to cryptography resistant to quantum computing, leaving the timetable to later government rules. Proposed investment and tax incentives would support security research, sovereign cloud services, secure data centres and cryptographic technologies, with assistance for smaller businesses and startups. The draft calls for avoiding policies that give a particular product or supplier a monopoly.

Quantum resistant cryptography is designed to withstand attacks using future quantum computers. The government's concern includes information being collected in encrypted form now and decoded later as computing capabilities advance.

The legislative timetable has shifted during drafting:

• August 2026: Public consultation closed, according to the earlier drafting schedule.
• September 7, 2026: An earlier description placed the draft under Ministry of Justice review.
• October 26, 2026: The government presented the proposal, with group discussions scheduled that afternoon.
• July 1, 2027: The parliamentary proposal's intended effective date, replacing the December 1, 2027, date described earlier.

The bill would amend the Cybersecurity Law and Data Law and preserve valid existing records and assessment results through transitional provisions. Existing systems would receive a route to meet the new requirements. Final obligations will depend on the text approved by the National Assembly and the implementing rules.

Key Points

• Vietnam proposes five data security risk levels, with protections increasing as potential harm rises.
• Two administrative approval procedures would cover system designs and overseas transfers involving levels 4 and 5.
• AI safeguards would address training data, unauthorized extraction and human oversight, with assessments integrated into AI procedures.
• Parliamentary reviewers seek tighter level 5 controls and stronger protection for biometric and genetic data.
• The revised proposal drops revenue based fines and a planned new national centre.
• The government proposes July 1, 2027, as the effective date, subject to parliamentary approval.