On 15 October 2024 Singapore’s Cyber Security Agency published the Guidelines on Securing AI Systems together with a Companion Guide. The pair proposes a secure-by-design approach applied across the full lifecycle of an AI system, from data sourcing and model development through deployment, operation and decommissioning.
Security as a lifecycle obligation
Most security regulation attaches at a moment in time: a certification before launch, a notification after an incident. Singapore’s guidelines instead treat security as continuous, on the reasoning that an AI system’s risk profile changes as its data, weights and operating environment change. This has an immediate documentary consequence — a single pre-launch test report will not satisfy an assessor looking for evidence of ongoing control.
The guidelines are also deliberately non-binding. That is a familiar Singaporean method: publish a method, let the market adopt it, and let adoption create the expectation. The later AI safety package continues in the same direction, as set out in our note on Singapore’s AI safety initiatives.
The Companion Guide is the more immediately useful half for engineering and legal teams, because it converts the principles into actions at each stage — for example, provenance checks when assembling training data and access controls around model weights. Where a jurisdiction offers both a principle document and an implementation companion, the companion is usually the document that ends up in procurement questionnaires.
Where it fits internationally
The lifecycle framing is converging across jurisdictions rather than diverging. The G7 cyber expert group statement on AI cybersecurity makes a similar argument at the political level, and the NIST GenAI pilot text evaluation works on the evaluation side. For Chinese companies operating across both markets, the useful comparison is with the domestic GB/T 45654 generative AI security standard, which also organises requirements by lifecycle stage. Governance structures for the same subject appear in our note on the AI safety governance committee.
What to implement now
Three actions carry most of the weight. First, produce an inventory of AI systems with their data sources, third-party models and interfaces, because lifecycle security is impossible to demonstrate without one. Second, extend the existing secure development lifecycle to cover model-specific steps: training data provenance, weight protection, prompt-injection testing and monitoring after release. Third, align the evidence you generate with more than one framework, since a Chinese security assessment and a Singapore-style lifecycle record overlap substantially and can be produced from the same underlying logs.
This article is based on the Cyber Security Agency of Singapore’s Guidelines and Companion Guide on Securing AI Systems and public information and is for reference only; it does not constitute legal advice.