Guidelines on Securing AI Systems
GUIDELINES ON SECURING AI SYSTEMS
The Cyber Security Agency of Singapore (CSA) has developed Guidelines on Securing AI Systems to help system owners secure the use of AI throughout its lifecycle. The document covers understanding AI threats, taking a lifecycle approach, starting with a risk assessment, and specific guidelines for planning and design, development, deployment, operations and maintenance, and end of life. These guidelines are non-mandatory and intended to be used alongside existing security best practices and the Companion Guide on Securing AI Systems.
Summary and content fields are drawn from the document text and are pending editorial review.
Who issued it
Who it governs
System owners and organizations adopting or considering the adoption of AI systems.
What the document contains
The document addresses cybersecurity risks to AI systems across five stages of the AI development lifecycle: planning and design, development, deployment, operations and maintenance, and end of life. It covers both classical cybersecurity threats and adversarial machine learning threats, alongside risk assessment steps and mitigation strategies.
The document does not address AI safety, fairness, transparency, inclusion, or the malicious misuse of AI in cyberattacks, mis/disinformation, and deepfakes. It also does not provide legally mandatory or prescriptive requirements.
- AI systems are vulnerable to both classical cybersecurity risks and novel adversarial machine learning attacks.
- System owners should adopt a secure-by-design and secure-by-default approach across the entire AI lifecycle.
- Conducting regular security risk assessments tailored to AI is foundational before implementing security controls.
- Supply chain components like training data, models, APIs, and libraries require continuous assessment and monitoring.
- Organizations should establish incident management, continuous monitoring, and vulnerability disclosure procedures for AI systems.
1. Introduction 1.1. Purpose and scope of this document 2. Understanding AI threats 3. Securing AI 3.1. Take a lifecycle approach 3.2. Start with a risk assessment 3.3. Guidelines for securing AI systems Glossary Annex A
How it scores
Each score is backed by a quote verified against the document text. Editorial review has not yet confirmed them.
Accountability
1/5Clear responsibility for AI outcomes
All stakeholders involved across the lifecycle of an AI system should seek to better understand the security threats and their potential impact
3.1. TAKE A LIFECYCLE APPROACH
screening score · 2026-10-23
Fairness
0/5Non-discrimination and equitable treatment
1.1. PURPOSE AND SCOPE OF THIS DOCUMENT
screening score · 2026-10-23
Human Dignity
0/5Respect for the inherent worth of persons
1.1. PURPOSE AND SCOPE OF THIS DOCUMENT
screening score · 2026-10-23
Human Oversight
0/5Humans remain in control of consequential decisions
1. INTRODUCTION
screening score · 2026-10-23
Privacy
2/5Protection of personal data and private life
Insecure development can also make AI models vulnerable to attacks (e.g. poisoning) that result in compromised model behaviour, or expose models and other intellectual property to theft, unauthorised replication or misuse.
2.4. Secure the AI development environment
screening score · 2026-10-23
Transparency
0/5Openness about how AI systems work and are used
1.1. PURPOSE AND SCOPE OF THIS DOCUMENT
screening score · 2026-10-23
Enforcement
0/5Consequences and compliance mechanisms with real force
1. INTRODUCTION
screening score · 2026-10-23
Impact Assessment
4/5Required assessment of an AI system's risks/impacts before or during use
Conduct a risk assessment, focusing on security risks related to AI systems, either based on best practices or your organisation’s existing Enterprise Risk Assessment/Management Framework.
3.2. START WITH A RISK ASSESSMENT
screening score · 2026-10-23