Enforce Input/Output Guardrails at the LLM Gateway layer — PII redaction, Prompt Injection defense, Jailbreak detection, Toxicity filter, and Tool Allow-list. Integrates Bedrock Guardrails, NeMo Guardrails, Llama Guard 3, and regex/regex-ML policies on Bifrost/LiteLLM with Langfuse audit trail.
Scanned 9/12/2026
Install to Claude Code
npx -y skills add aibot88/sec_skill_store --skill ai-gateway-guardrails --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Ai Gateway Guardrails?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/aibot88-ai-gateway-guardrails)More formats (shields.io, HTML) on the badges page.
---
name: ai-gateway-guardrails
description: "Enforce Input/Output Guardrails at the LLM Gateway layer — PII redaction, Prompt Injection defense, Jailbreak detection, Toxicity filter, and Tool Allow-list. Integrates Bedrock Guardrails, NeMo Guardrails, Llama Guard 3, and regex/regex-ML policies on Bifrost/LiteLLM with Langfuse audit trail."
argument-hint: "[compliance scope — ISMS-P, finance, healthcare]"
user-invocable: true
model: claude-sonnet-4-6
allowed-tools: "Read,Write,Edit,Bash,Grep,Glob,mcp__eks,mcp__aws-documentation,mcp__well-architected-security"
---
## When to Use
- 한국 금융권(전자금융감독규정·ISMS-P), 의료, 공공 등 규제 환경에 LLM 서비스를 배포할 때
- Prompt Injection / Jailbreak / PII 유출 / Tool Poisoning 위협을 방어해야 할 때
- Bedrock Guardrails, NeMo Guardrails, Llama Guard 3 중 선택 및 조합이 필요할 때
- Agent 가 외부 Tool 을 호출할 때 Allow-list 기반 정책이 필요할 때
## When NOT to Use
- 내부 PoC 로 위협 모델이 불필요 — Guardrail 오버헤드만 발생
- Bedrock 매니지드 모델만 호출하며 Bedrock Guardrails 기본 활성 — 추가 구성 불필요 (단, 로그는 필수)
- 내부 RAG 없이 단순 Q&A — regex 수준의 Input Guard 만 필요할 수 있음
## Preconditions
- Inference Gateway (Bifrost/LiteLLM) 가 이미 배포됨 (`inference-gateway-routing` 완료)
- Langfuse 가 audit log 를 수신 가능 (`langfuse-observability` 완료)
- PII 정책·차단 카테고리·Tool Allow-list 정의 문서 확보
## Procedure
### Step 1. 위협 모델 정의 (OWASP LLM Top 10 기반)
- LLM01 Prompt Injection (Direct/Indirect)
- LLM02 Sensitive Information Disclosure (PII, 영업비밀)
- LLM06 Excessive Agency (Tool 오용)
- LLM08 Vector & Embedding Weaknesses (RAG poisoning)
### Step 2. 다층 방어 (Defense in Depth)
```
User → Input Guard → Gateway Policy → Tool Allow-list → LLM → Output Guard → Response
↓
Audit Log (Langfuse)
```
- Input Guard: PII redaction, Injection pattern, Jailbreak classifier
- Gateway Policy: AuthN/Z, Rate Limit, Tenant Isolation
- Tool Allow-list: MCP Server Registry, Scoped tokens
- Output Guard: PII scrub, Toxicity, Fact check
- Audit Log: 모든 단계에서 Langfuse + CloudTrail 기록
### Step 3. Bedrock Guardrails 연동 (매니지드)
```yaml
# Bifrost 설정
providers:
bedrock:
region: ap-northeast-2
guardrails:
- id: arn:aws:bedrock:ap-northeast-2:ACCOUNT:guardrail/PII-BLOCK
version: "1"
- id: arn:aws:bedrock:ap-northeast-2:ACCOUNT:guardrail/TOXICITY
version: "1"
```
### Step 4. NeMo Guardrails (오픈소스 Flow)
```yaml
# config.yml
models:
- type: main
engine: openai
model: gpt-4.1
rails:
input:
flows:
- self check input
- detect pii
output:
flows:
- self check output
- remove pii
- fact checking
```
### Step 5. Llama Guard 3 (Output Classifier)
- Meta Llama Guard 3 8B 모델을 vLLM 별도 Pod 로 배포
- Bifrost output 훅에서 Llama Guard 3 call → unsafe 판정 시 재생성 또는 차단
### Step 6. Tool Allow-list (MCP)
```yaml
mcpAllowList:
- name: aws-documentation
scopes: ["read"]
- name: eks
scopes: ["read", "describe"]
# deny all others
tokenPolicy:
maxLifetimeSeconds: 900
audience: agentic-platform
```
### Step 7. Audit & 알림
- 모든 guard violation 은 Langfuse `scores` + `tags` 로 기록
- Prometheus 메트릭 `guardrail_violation_total{type="pii",decision="block"}`
- CloudWatch Logs + SIEM 연계 (Security Lake)
- Slack/PagerDuty 알림 기준: `guardrail_violation_rate > 5%/5m`
## Good Examples
- ISMS-P 대상 금융: Bedrock Guardrails(managed PII + Block) + NeMo Guardrails(자체 Policy) + Llama Guard 3(output)
- Coding Agent: Tool Allow-list 로 `shell_exec`, `network_request` 차단
- RAG: Indirect Injection 방어용 Llama Guard 3 + fact-check Flow
## Bad Examples (금지)
- Guardrails 없이 Tool-calling Agent 를 프로덕션 배포 → LLM06 즉시 위반
- 정규식 기반 PII 단독 → 한국 주민번호 변형 패턴 미탐지, ML classifier 병행 필수
- Audit log 미수집 → 규제 감사 시 근거 부재
- `allowed-tools: ["*"]` — 전체 허용 = 정책 없음
## References
- AI Gateway Guardrails (community resource)
- 컴플라이언스 프레임워크 (community resource)
- [Bedrock Guardrails 공식 문서](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html)
- [NeMo Guardrails](https://github.com/NVIDIA/NeMo-Guardrails)
- [Llama Guard 3 (Hugging Face)](https://huggingface.co/meta-llama/Llama-Guard-3-8B)
- [OWASP LLM Top 10 2025](https://genai.owasp.org/llm-top-10/)
- [ISMS-P 인증 기준](https://isms.kisa.or.kr/) — 한국 인터넷진흥원
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!