Generative AI assistants, retrieval-augmented generation (RAG) systems, and AI agents create new attack surfaces, from prompt injection to data leakage through model outputs. AI security is a natural extension of WC1’s AI Solutions work: we secure the AI systems we and others build, and help organisations in Singapore and the Philippines govern how staff use third-party AI tools.
The problem: why AI security matters
LLM applications can be manipulated by crafted inputs, including instructions hidden in documents or web pages they retrieve. Agents with tool access can be tricked into taking actions or exposing data. Sensitive data can leak through prompts, logs, or model outputs.
At the same time, staff adopt public AI tools faster than policies can keep up, and boards ask whether AI use is safe and compliant. Traditional security testing does not cover these AI-specific risks.
Who this service is for
Organisations deploying copilots, chatbots, RAG systems, or AI agents; regulated entities introducing AI into customer-facing or decision-making processes; and leadership teams setting AI usage policy.
What WC1 delivers
Every engagement is scoped to your environment and risk. These are the building blocks we combine for AI security:
- AI threat modelling: Assessment of AI applications against the OWASP Top 10 for LLM Applications and MITRE ATLAS, covering data flows, tools, and trust boundaries.
- Prompt injection and output controls: Input and output filtering, privilege separation for agents, and grounding controls that reduce manipulation and hallucination impact.
- Data protection for AI: Controls over training data, retrieval sources, logging, and retention, aligned to data protection law.
- AI usage governance: Policies, approved tool lists, and technical controls for staff use of public and enterprise AI tools.
- AI red teaming: Adversarial testing of AI applications to find jailbreaks, data leakage, and unsafe agent behaviour before attackers do.
- Model and AI supply chain security: Review of third-party models, plugins, and AI vendors, including provenance and contractual protections.
How the engagement runs
1. Inventory
Identify AI applications, tools in use, data sources, and owners.
2. Assess
Threat model priority applications and review governance and data controls.
3. Harden
Implement technical controls, guardrails, and monitoring for AI applications.
4. Test
Run AI red teaming and fix findings.
5. Govern
Establish AI risk review, usage policy, and ongoing monitoring.
Commercial shape follows the work: a fixed-scope assessment or implementation when the problem is well bounded, or a retainer when you need ongoing operation, tuning, and reporting. Coverage hours, response times, and responsibilities are written into the statement of work so there are no surprises.
Deliverables you can expect
- AI application inventory and risk register
- Threat models for priority AI applications
- Guardrail and monitoring configuration
- AI usage policy and approved tool framework
- AI red team report and remediation plan
- AI vendor and model due diligence checklist
Platforms and tools we integrate
We work with Palo Alto Networks Prisma AIRS and AI Access Security, Microsoft Purview and Defender for Cloud AI threat protection, Azure AI Content Safety, AWS Bedrock Guardrails, Google Cloud Model Armor, and open-source tools such as garak and PyRIT for AI red teaming, alongside the AI platforms WC1 builds on.
WC1 is vendor-neutral: we integrate the platforms you already own before recommending anything new. Naming a product here describes technology we work with; it does not imply a partnership, reseller agreement, or vendor certification.
Singapore: regulatory and market relevance
Singapore’s Model AI Governance Framework for Generative AI, IMDA’s AI Verify toolkit, and CSA’s Guidelines on Securing AI Systems provide the local reference points for AI security. MAS has issued guidance on AI risk management for financial institutions, and the PDPC’s Advisory Guidelines on use of personal data in AI recommendation and decision systems apply to personal data. We map controls to these frameworks.
Philippines: regulatory and market relevance
The Philippines’ NPC has issued guidance on the application of the Data Privacy Act to AI systems processing personal data, and DICT is developing national AI governance initiatives. BSP expects supervised financial institutions to manage risks of new technologies within their IT risk management frameworks under Circular 1140. Philippine BPO and shared services firms using AI on client data face additional contractual scrutiny.
How this fits the rest of your security programme
AI Security rarely stands alone. It works best alongside Data Security and DLP, Application and API Security, Identity Security and Access Management, Cybersecurity Compliance and Advisory. It also connects to our AI Solutions, Private AI in Singapore services. See the full catalogue on our Cybersecurity services page.
Related reading: Prompt injection security checklist for enterprise AI; LLM output logging and retention under PDPA.
Frequently asked questions
What is prompt injection?
An attack where crafted input causes an AI model to ignore its instructions, reveal data, or take unintended actions. It can come directly from a user or indirectly from content the model retrieves.
Can prompt injection be fully prevented?
Not completely with current technology. The practical approach is defence in depth: limit what the model can access and do, filter inputs and outputs, and monitor for abuse.
Should we ban public AI tools?
Outright bans are hard to enforce and push use underground. Most organisations do better with approved tools, clear policy, and DLP controls on sensitive data.
Do you secure AI systems built by other vendors?
Yes. We assess and test AI applications regardless of who built them, and review vendor controls and contracts.
How does this relate to your AI Solutions services?
Our AI Solutions pillar builds governed AI products; AI security assures them. Security reviews are part of our AI delivery, and this service extends that to any AI system.
Next step
Tell us which AI applications you run or plan to launch. We will propose a threat modelling session and a security review plan. Contact WC1 for a no-obligation conversation; we will tell you honestly if we are not the right fit.