Singapore's CSA publishes Guidelines on Securing AI Systems

Singapore's Cyber Security Agency published its Guidelines on Securing AI Systems alongside a Companion Guide on 15 October 2024.

The guidelines give system owners a document that aims to support system owners adopting or considering AI by identifying potential security risks and setting out guidance to mitigate them across the AI lifecycle.

What it is

The guidelines are non-mandatory but strongly encouraged for organisations building or deploying AI in Singapore.

They are explicitly scoped to cybersecurity risk and do not try to cover every AI concern.

Alongside the guidelines, CSA worked with practitioners to produce a companion document of practical controls, developed by collaborating with AI and cybersecurity practitioners to build a Companion Guide on Securing AI Systems.

What it requires

The guidance is structured around a lifecycle approach comprising five stages: Planning and Design, Development, Deployment, Operations and Maintenance, and End of Life.

Underneath that structure sits a four-step risk assessment process underpinning the entire framework, directing organisations to identify risks, prioritise them by impact and available resources, implement relevant controls, and evaluate residual risks.

The aim is to protect AI systems against classical cybersecurity risks such as supply chain attacks, and novel risks such as Adversarial Machine Learning.

To help organisations self-check progress, the guidelines also include a self-assessment checklist and a glossary of terms to help organisations assess and improve their AI security posture and awareness.

Feedback gathered during consultation helped refine the guidelines and align them with international standards such as those from the U.S. Cybersecurity and Infrastructure Security Agency and the UK National Cyber Security Centre.

What security leaders should do

Security leaders should map their own AI projects against the five lifecycle stages rather than treating model hardening as sufficient on its own, since the guidance frames model-only hardening as incomplete. Run the four-step risk assessment early in planning and design so that supply chain, data, and adversarial machine learning risks are prioritised before deployment decisions are locked in. Use the self-assessment checklist to benchmark current controls, and treat the Companion Guide as a living reference to update playbooks as new practical measures are added. Because the guidelines stop short of covering AI safety, fairness, or malicious use of AI for deepfakes and scams, leaders should pair this framework with separate internal policies addressing those adjacent risks.