AI Security Risks Philippines: Prompt Injection & Data Leakage

Last materially reviewed: September 4, 2026

Direct Answer

Generative AI introduces security risks that are different from ordinary SaaS use. Philippine businesses should address prompt injection, sensitive information disclosure, insecure integrations, supply-chain weaknesses, excessive model agency, unsafe output handling and shadow AI. The practical control set is to approve tools, restrict sensitive inputs, isolate high-risk integrations, apply least privilege, validate outputs, log important actions, review vendors and require human approval before consequential actions.

Key Takeaways

Prompt injection can manipulate an AI system through direct prompts or untrusted external content.
Sensitive information can leak through prompts, retrieval systems, training data, logs or generated output.
Shadow AI increases risk because employees may use tools outside approved security and privacy controls.
AI agents and connected tools should have tightly limited permissions and human approval for high-impact actions.
AI security should be included in vendor reviews, incident response, employee training and data-loss prevention controls.

1. Prompt Injection

OWASP identifies prompt injection as a leading LLM application risk. A malicious or unintended instruction can influence model behavior, including through files, webpages, retrieved documents or connected systems. Do not assume a system prompt alone can prevent this risk.

2. Sensitive Information Disclosure

Employees may enter customer, employee, financial, authentication or confidential business information into an AI tool. Models and applications may also expose data through outputs, logs, retrieval systems or integrations. Use data classification, prompt restrictions, approved tools, access controls and retention settings.

3. Shadow AI

Shadow AI refers to unapproved AI tools or workflows introduced outside formal governance. The risk is not solved by a blanket ban alone. Businesses should provide approved alternatives, define prohibited data, train staff, monitor risky integrations and establish a simple process for requesting new tools.

4. Excessive Agency and Connected Actions

AI systems that can send email, modify records, execute code, approve transactions or call APIs can create higher operational risk. Give each system the minimum permissions needed and require human approval for sensitive or irreversible actions.

5. Supply-Chain and Vendor Risk

AI products may depend on external models, plugins, datasets, APIs and subprocessors. Review the full processing and dependency chain where feasible, including update practices, incident handling and how vendor changes affect the business workflow.

6. Improper Output Handling

Generated output should be treated as untrusted input when it is passed to code, databases, browsers, workflows or downstream systems. Validate, sanitize and constrain outputs before allowing them to trigger actions.

AI Security Control Checklist

Maintain an approved AI tool list.
Block or restrict sensitive data from unapproved tools.
Use enterprise authentication and least privilege.
Separate test data from production data.
Review integrations, plugins and autonomous actions.
Validate generated output before execution or publication.
Log consequential actions and incidents.
Train staff on prompt injection, data leakage and shadow AI.
Include AI systems in incident response and vendor reviews.

Decision Snapshot

The more an AI system can see, retrieve or do, the more it should be treated like privileged software rather than a chatbot.

AI Security Risks for Philippine Businesses: Prompt Injection, Data Leakage and Shadow AI