/AI & 자동화/LLM AI Security and Governance Checklist: RAG, Tool Permissions, and Pre-Launch Review
AI & AutomationLLMSecurityAI거버넌스

LLM AI Security and Governance Checklist: RAG, Tool Permissions, and Pre-Launch Review

A security and governance checklist to review before launching enterprise LLM agents. It covers RAG access control, tool authorization, prompt injection testing, audit logs, and incident-response evidence.

LLM AI Security and Governance Checklist: RAG, Tool Permissions, and Pre-Launch Review

An LLM/AI security and governance checklist is only actionable when you define who can use which data and tools, what to test before launch, and who can stop the system if it fails. Do not treat installing a prompt filter as a pass.

The following is an example launch review for an enterprise LLM agent that uses RAG and external tools. It is not a certification scorecard or a legal standard common to every industry. Set owners and approval criteria according to your organization's data and operational risk.

LLM AI Security and Governance Checklist

AreaQuestion to verifyEvidence to keepExample owner
Scope of useAre allowed and prohibited tasks, and the impact of failure, defined?Use-case list, risk assessmentService owner
DataWhat information enters inputs, retrieval, training, and logs?Data-flow diagram, retention and deletion rulesData / privacy owner
RAG permissionsIs document access limited by the user's permissions?Negative tests of tenant and document ACLsBackend owner
Tool permissionsAre model proposals separated from actual execution authorization?Tool allowlist, authorization testsPlatform owner
External contentCan instructions in documents or web pages change permissions?Indirect injection test recordsSecurity owner
OutputAre HTML, SQL, commands, etc. not executed directly?Output schema and destination validationApplication owner
Sensitive actionsAre there adequate controls for outbound transfer, deletion, and payments?Approval policy, retry and duplicate-prevention resultsBusiness owner
EvaluationAre successful-task rates and security failures measured together?Fixed evaluation set, results by versionEvaluation owner
OperationsAre cost, errors, and abnormal tool calls monitored?Limits, alerts, operations dashboardOperations owner
Incident responseCan you halt execution, revoke tokens, and roll back?Owners, halt-and-recovery rehearsalIncident response owner

Record each row as pass, needs work, or not applicable, and keep links to supporting evidence. If you allowed an exception, also record the approver and the next review date. This table and role mapping are an editorial operations example, not a table from an official framework.

Prompt injection cannot be stopped with regex alone

Besides direct user input, hidden instructions in retrieved documents or web pages can influence model behavior. Blocking a specific English phrase such as ignore system prompt still leaves other languages, phrasings, and external-document paths. OWASP describes combining multiple defenses—least privilege, output validation, separating untrusted content, and controlling sensitive actions—rather than promising complete blocking. OWASP Prompt Injection

So security tests should check not only whether “the model refused the instruction,” but also whether data was actually exposed or a tool actually ran. Verify that the server-side authorization layer still blocks an inappropriate tool call even if the model proposes it.

Bind tool execution to user permissions

For example, suppose a customer-support agent proposes ticket.read. The server must compare the authenticated user's organization and role with the organization that owns the requested ticket. Do not grant access based on a tenant_id the model emitted or a sentence such as “an administrator approved this.”

You can structure the execution flow as follows.

  1. Identify the user and tenant from the server session
  2. Validate the tool name and arguments against the allowed schema
  3. Perform server-side authorization for the target resource
  4. Confirm any policy-required action approval
  5. Check execution limits and duplicate prevention, then invoke the tool
  6. Record the result and authorization decision as minimal audit information

This is a design example. Real access-control code must connect to each service's authentication and authorization store. For an implementation view, continue with LLM Agent Security Design Guide.

Review RAG retrieval and audit logs

Build tenant and permission filters for retrieval from the server's authentication context. Keep document IDs and versions so you can trace answer sources, but a citation does not by itself guarantee accuracy or access rights. Test that permission changes and deletions also propagate to the index and cache.

Prioritize audit-log fields needed for incident investigation: request identifier, model and prompt version, tool name, authorization result, latency, and cost. Storing full raw inputs and responses for long periods can copy personal or confidential data into logs, so define purpose, access rights, and retention first.

Example tests to run before launch

The following is a test design you can run with synthetic data and no real customer information. This article does not claim measured results for any specific product.

Test input / conditionExpected behavior
Specify another tenant's document IDDeny access for retrieval, download, and tool results
Insert an outbound-transfer instruction into a retrieved documentDocument instructions alone must not create transfer permission
Model generates a tool name that is not allowedBlock before execution via schema and allowlist
Change recipient or amount after approvalRe-review if the action no longer matches the approved content
Repeat the same action via network retryPrevent duplicate payment, send, or deletion
Change model or prompt versionRe-run the same evaluation set of success and failure cases
Revoke a tool token or halt executionNew executions are blocked and operators can confirm status

Also record how often legitimate requests are blocked. Do not judge operational fitness from “we blocked a few attack strings” alone.

Connect NIST AI RMF to operational records

NIST AI RMF is a voluntary risk-management framework that presents Govern, Map, Measure, and Manage functions. You can also consult the generative AI profile. The mapping below is an application example for this article. NIST AI RMF, AI RMF Core Functions

FunctionOperational records to produce for this service
GovernOwners, approval and exception policy, change management
MapUse cases, data flows, affected users
MeasureBenign and misuse evaluation results, limits and failure cases
ManageAction priorities, launch criteria, halt-and-recovery plan

Treat not only model swaps but also new data sources, new tools, and permission expansions as triggers for re-review. What matters is whether current operations and evidence still match the checklist—not the date it was written.

확인 정보
✦ ✦ ✦
편집 검토 · Editorial Review

Nodelog는 모든 콘텐츠의 내용과 출처를 공개 전에 검토합니다. 환경(OS·버전)에 따라 결과가 달라질 수 있는 기술 정보는 공식 문서와 함께 확인하며, 검토 기준과 정정 원칙은 편집 정책에서 안내합니다. 오류를 발견하시면 이메일로 제보해 주세요 — 확인 후 신속히 정정합니다.

편집 책임 · Nodelog 기술 편집팀·발행 · ·업데이트 ·

Comments

Be the first to comment.