AI Guide For Boards in Singapore

AI GUIDE FOR BOARDS IN SINGAPORE

Section V: Organisational Resilience and Capability

Oversight of AI Infrastructure

CAPACITY PLANNING

HIGH AVAILABILITY AND REDUNDANCY

PERFORMANCE UNDER STRESS

1

2

3

Capacity planning should ensure infrastructure can support peak AI demand without disrupting core business operations, with sufficient computational capacity across GPUs, storage, and network bandwidth to support AI workloads at scale.

High availability and redundancy help ensure AI-enabled services remain available even if disruptions or failures occur, through failover capabilities, geographic redundancy, and disaster recovery arrangements.

Agentic AI systems that interact with external

systems, APIs, and real-time data streams may introduce latency and operational failure points. Stress testing

before deployment is therefore important.

10.2 Cyber Security Resilience

• Model the: Unauthorised extraction or replication of proprietary models and intellectual property. • Prompt injection: Malicious prompts designed to bypass controls, manipulate outputs, or expose sensitive information.

Boards should view AI cyber security with a broad lens. e focus is not limited to protecting models themselves, but extends to safeguarding enterprise data, connected systems, access pathways and AI-enabled business processes. In many organisations, immediate risks involve sensitive data exposure, insecure integrations, misuse of privileged access, or weak governance over how AI tools are deployed and used.

10.2.2 Ensuring strong controls

Secure AI Deployment depends on strong controls, including encryption, identity and access management, administrative safeguards and privacy protections. ese controls are important not only for compliance, but also for maintaining trust in AI-enabled systems. Key governance considerations include: • Zero-trust architecture: Applying least-privilege access, encryption and segmentation of AI workloads from core enterprise systems. • Endpoint protection: Securing inference application programming interface through authentication, monitoring, rate limiting and logging.

10.2.1 Evolving threat landscape

AI introduces new categories of cyber threats, including: • Model poisoning: Malicious manipulation of training data to inuence model behaviour or outputs. • Data exltration: e or exposure of sensitive data used within AI pipelines. • Adversarial attacks: Carefully craed inputs designed to mislead or manipulate AI systems.

91

Made with FlippingBook Digital Proposal Creator