Draft Data Security Law Would Require Vietnam’s AI Developers to Vet Training Data and Guard Against Poisoning
Vietnam’s draft Law on Data Security, presented to the National Assembly on October 26 by Minister of Public Security General Luong Tam Quang, includes a set of obligations aimed directly at organisations that build and deploy artificial intelligence, according to the ministry and the newspaper Báo Công Thương. The draft has six chapters and 29 articles, places data under five risk levels and gives the Ministry of Public Security the lead role in state management.
Training and test data would have to come from legitimate sources, and developers must control its origin and integrity and take steps against data poisoning. They would also need technical measures to stop unauthorised extraction, reverse engineering or reconstruction of training data through queries, along with access controls and security logs. Systems handling Level 3 data or above must build protection in from the design stage, and Level 4 and 5 systems would need security testing before deployment.
The draft treats derived data, including model weights and algorithm parameters, as protected. For AI agents that can plan or act without approval, it requires unique system identification, limited permissions and logging, and humans must be able to intervene, pause or revoke access. If a poisoned system produces output that threatens national security, the operator would have 72 hours to fix it before emergency technical measures apply.
What it means in practice
To avoid duplicate paperwork, data security assessments for AI systems would be folded into the risk assessment already required under the AI law, and the ministry would take part in controlled testing led by the AI regulator. That links the bill to the generative AI sandbox, which officials recently clarified does not grant deployment rights. Transfers abroad of Level 1 and 2 data would need no procedure, while only Level 5 data critical to national defence would be barred from leaving the country.
The figures presented to deputies show why the government is moving: about 552,000 cyberattacks hit Vietnamese information systems in 2025, and authorities uncovered around 300 cases involving the sale of more than 1.7 million data records. As Vietnam courts foreign partners, including through its proposed ASEAN-China AI pact, startups will want clarity on what counts as a legitimate source of training data.
The law is expected to take effect on July 1, 2027, with existing systems given 24 months to comply and small and medium-sized businesses 36 months. What to watch is how deputies amend the AI provisions during debate in the current session.