Tech BlogNovember 18, 2026Sarah Kim1 views

Generative AI Utilization Guidelines and Security: A Checklist for In-house AI Adoption

The adoption of in-house generative AI entails various risks such as information leakage, prompt injection, inaccurate responses, and copyright and personal information infringement. Therefore, organizations must establish clear utilization guidelines and a thorough security checklist to prevent these issues. This guide encompasses major domestic and international directives and technical control measures, presenting practical strategies for a secure AI Transformation (AX).

#Generative AI Guidelines#AI Utilization Guidelines#AI Security#In-house AI#LLM Security#AI Adoption
Generative AI Utilization Guidelines and Security: A Checklist for In-house AI Adoption
Sarah Kim

November 18, 2026

Key Answer: The adoption of in-house generative AI presents a significant AI Transformation (AX) opportunity for enhancing work efficiency. However, it can entail severe security and legal risks, including information leakage, prompt injection, inaccurate responses, and infringement of copyright and personal information. Organizations must establish clear utilization guidelines and a thorough security checklist to minimize these risks and safely realize AI's potential.

Why In-house Generative AI Guidelines are Necessary

Generative AI technology is accelerating AI adoption (AI Transformation, AX) within organizations, offering enhanced productivity and new business opportunities. However, utilizing generative AI internally without clear guidelines can lead to various unforeseen security and operational risks.

First, there is the risk of information leakage. When documents containing internal information are input into a generative AI model to increase work efficiency, there is a possibility that such data could be transmitted to external AI model providers, leading to leakage or malicious use for model training. The unintended exposure of sensitive internal data or confidential information can result in significant losses for organizations.

Second, there is the risk of Prompt Injection attacks. Prompt injection attacks, where attackers deliberately inject malicious prompts to hijack system control or induce abnormal responses, are a specific threat to LLM-based systems. This new attack vector is difficult to defend against with traditional web application security measures and can lead to sensitive information leakage or system malfunctions.

Third, issues arise due to inaccurate responses (hallucinations). Generative AI can sometimes produce plausible yet inaccurate responses (hallucinations) that deviate from facts. Utilizing such information directly in operations can lead to flawed decision-making or a decline in trustworthiness. Especially in areas requiring professional judgment, such as law, health, and finance, thorough verification of AI responses is essential.

Fourth, there is the possibility of copyright and personal information infringement. AI-generated output may include parts of existing copyrighted works or output personal information from training data in an identifiable format, leading to issues of copyright and personal information infringement. This can result in legal disputes and sanctions from regulatory authorities, undermining an organization's social responsibility.

Summary Table of Major Domestic Guidelines

Major domestic institutions have published various guidelines for the safe and responsible utilization of generative AI. Organizations must thoroughly review and comply with these directives when formulating their AI adoption (AX) strategies.

Document TitleInstitutionPublication DateKey Content
Security Guidelines for Utilizing Generative AI such as ChatGPTNational Intelligence ServiceJune 2023Prohibition of inputting internal/personal information, verification of responses, use of official plugins/extensions, API key protection, and recommendation for establishing organization-wide utilization policies.
AI Security Guidebook for National and Public InstitutionsNational Intelligence Service, National AI Security CenterDecember 2025AI system security threats and countermeasures, security measures for agentic and physical AI systems, lifecycle-specific threats and responses, security checklist, and recommendations for major generative AI service security settings. References NIST AI RMF, ISO/IEC 42001, OWASP LLM Top 10, MITRE ATLAS.
Public Sector Hyperscale AI Adoption and Utilization Guidelines 2.0National Information Society Agency (NIA) · Digital Platform Government CommitteeApril 2025 (1.0 revised in April 2024)Public AI adoption direction, adoption procedures and considerations, performance management, accompanied by case studies.
Public Sector AI Adoption and Utilization GuideMinistry of Interior and SafetyJune 2026Guidance on 5 stages of service establishment (Planning-Budget-Contract-Construction-Operation), RAG-first strategy for hallucination suppression, utilization of government-wide AI common platform.
Guidelines (regarding non-public administrative information and personal information)Ministry of Interior and SafetyFebruary 2025Do not input non-public administrative information or personal information into AI services, and use AI outputs only after sufficient verification.
Guide to Personal Information Processing for the Development and Utilization of Generative Artificial Intelligence (AI)Personal Information Protection CommissionAugust 2025Three types: utilizing LLM as a Service (API), fine-tuning existing/open-source LLMs, and building proprietary SLMs; minimum safety measures for each of 4 stages; data contamination/jailbreaking response; AI agent management; CPO-centric governance.
Revision of Financial Sector Artificial Intelligence GuidelinesFinancial Services CommissionAnnounced June 18, 2026, enforced June 22Self-regulation for all financial companies utilizing AI, based on 7 principles (Governance, Legality, Subsidiarity, Reliability, Financial Stability, Good Faith, Security). Accompanied by the Financial Supervisory Service's "Financial Sector AI Risk Management Framework" and the Financial Security Institute's "Financial Sector Artificial Intelligence Security Guide."
Financial Sector Network Segregation Improvement RoadmapFinancial Services CommissionAugust 2024Conditional allowance of generative AI utilization through innovative financial services (sandbox), subject to security conditions.

Key Checkpoints for Organizations Following the Enforcement of the AI Basic Act

The "Framework Act on the Promotion of Artificial Intelligence and the Establishment of a Trustworthy Foundation," scheduled to take effect on January 22, 2026, significantly impacts organizations' AI adoption (AX) strategies. Specifically, organizations must thoroughly review and prepare for the following provisions.

  • Article 31 Transparency: Organizations must provide prior notification that products/services are based on high-impact AI and generative AI, indicate that outputs are generated by AI, and ensure deepfake-like outputs are clearly identifiable.
  • Article 34 Obligations of High-Impact AI Business Operators: High-impact AI business operators must establish risk management plans, provide explanations of results to the technologically feasible extent, implement user protection measures, ensure human oversight and supervision, and retain documentation of safety and reliability measures.
  • Fines and Grace Period: Article 43 includes provisions for fines, and the government has announced a grace period of at least one year for fines. The primary subjects of obligation under this law are AI business operators (not end-users).
  • Guideline Publication: The Ministry of Science and ICT plans to publish the "Guidelines for Ensuring AI Transparency" in January 2026.

In-house Generative AI Utilization Guidelines Table of Contents Template

Establishing clear and systematic utilization guidelines is essential for successful in-house AI adoption (AX). The following is a template for an in-house generative AI utilization guideline table of contents that organizations can adopt.

  1. Introduction and Purpose
  2. Fundamental Principles for Generative AI Utilization
  3. Secure Generative AI Utilization Methods (Sensitive Information, Prompts, Output Verification, Copyright, Personal Information)
  4. Internal AI System Usage Rules (Approved Tools, Access Management, Log Management)
  5. Roles and Responsibilities (Governance, Responsible Personnel, Training)
  6. Measures and Sanctions in Case of Violation
  7. Appendix (Relevant Laws and Regulations, FAQ)

Security Checklist

Generative AI security, which is essential for an organization's AI adoption (AX), requires a multifaceted approach. Organizations are encouraged to assess the security level of their in-house AI systems using the following checklist.

Security ItemVerification Question
Input Control (Sensitive Information Masking)Has a system been established to automatically identify, mask, or remove sensitive information (e.g., personally identifiable information, trade secrets) included in user input prompts?
Prompt Injection DefenseIs a defense system in place to detect and block attempts by malicious prompts to bypass or control model behavior?
Output Verification and Source AttributionHas a procedure been established to verify that AI model generated outputs are factual and do not contain harmful or biased content, and are sources of information attributed when necessary?
Access Control, Logging, and AuditingAre access permissions to AI models and systems restricted to the minimum necessary, are detailed logs of all user activities and model inputs/outputs recorded, and are they regularly audited?
Model and Data Location (BYOM, On-premise Deployment)When processing sensitive data, are organizations considering reducing reliance on external LLM services and adopting a Bring Your Own Model (BYOM) strategy or using on-premise deployed models?
Human-in-the-Loop (HITL) ReviewHas a process been established to ensure that critical or sensitive AI-generated outputs undergo human review and approval before final determination?

Technical Control Implementation Methods (Guardrail Layer)

Establishing a guardrail layer through technical controls is essential for the safe utilization of generative AI. This guardrail monitors and controls inputs (prompts) flowing into the AI system and outputs generated externally in real-time, mitigating potential risks. For example, SeekersLab's KYRA AI Guardrail serves as a security gateway for enterprise LLM usage, providing prompt injection detection, LLM output validation, and adversarial attack simulation capabilities. This supports the secure utilization of generative AI in enterprise environments without the risk of sensitive data leakage. Further details on LLM guardrail establishment can be found in AX Part 4: Building an LLM Guardrail.

International Reference Standards

Major global institutions also propose various standards and frameworks to enhance AI trustworthiness and security. When adopting AI (AX), it is crucial for domestic organizations to reference international standards to advance their security frameworks.

  • OWASP Top 10 for LLM Applications 2025: Published in November 2024, this guideline presents 10 major security vulnerabilities specific to LLM-based applications. System Prompt Leakage and Vector and Embedding Weaknesses have been added.
  • NIST AI RMF 1.0 and Generative AI Profile (NIST AI 600-1): The National Institute of Standards and Technology (NIST) AI Risk Management Framework (AI RMF) provides methods for identifying, evaluating, and managing AI risks through four core functions: Govern, Map, Measure, and Manage. A generative AI-specific profile (NIST AI 600-1) was additionally published in July 2024.
  • ISO/IEC 42001:2023 AI Management System Standard: Established in 2023, this international standard specifies requirements for establishing, implementing, maintaining, and continuously improving an Artificial Intelligence Management System (AIMS) for organizations to responsibly develop and operate AI systems. It has a structure similar to ISO 27001 and allows for certification related to AI governance and risk management.

FAQ

Q1. Why are in-house generative AI guidelines necessary?

A1. In-house generative AI guidelines are necessary to prevent potential risks such as information leakage, prompt injection, inaccurate responses (hallucinations), and copyright and personal information infringement. They establish clear standards for the safe and ethical utilization of AI, enabling organizations to fulfill their legal and ethical responsibilities.

Q2. What is Prompt Injection and how can it be defended against?

A2. Prompt injection refers to attempts by attackers to alter an AI model's intended behavior or leak sensitive information through malicious prompts. To defend against this, technical controls are necessary, including input prompt validation, sensitive information masking, and the detection and blocking of malicious prompts through AI guardrail solutions such as KYRA AI Guardrail.

Q3. What is the impact of the AI Basic Act on an organization's AI adoption (AX)?

A3. The AI Basic Act specifies obligations such as transparency notification for high-impact AI and generative AI-based services, marking of generated outputs, and risk management and user protection duties for high-impact AI business operators. Organizations must review and incorporate relevant regulations throughout the entire AI system development and operation lifecycle to comply with these legal requirements.

Q4. How should organizations respond to the issue of generative AI producing inaccurate responses (hallucinations)?

A4. As AI hallucination is a persistent issue, organizations must establish processes to consistently verify AI outputs rather than blindly trusting them. Especially for critical decision-making or areas requiring professional judgment, it is necessary to implement a Human-in-the-Loop (HITL) approach to ensure final human review and to clearly attribute the sources of responses.

Q5. Which security checklist items should be prioritized when adopting in-house AI?

A5. When adopting in-house AI, priority should be given to input control (sensitive information masking), prompt injection defense, output verification and source attribution, AI system access control and log auditing, and model and data location strategies (BYOM, on-premise deployment) for sensitive data processing. Additionally, a Human-in-the-Loop (HITL) review process for critical outputs is also important.

Related Articles

Stay Updated

Get the latest security insights delivered to your inbox.

Tags

#Generative AI Guidelines#AI Utilization Guidelines#AI Security#In-house AI#LLM Security#AI Adoption