Can it detect if an LLM hallucinates an ID?

Yes, the output validation phase includes checks for UUID patterns in the response to ensure the model hasn't invented identifiers that weren't in the original context.

How does it protect against prompt injection?

It implements prompt auditing using regex patterns to detect forbidden identifiers and ensures a strict separation between system context and user-provided content.

Does this work with Retrieval Augmented Generation (RAG)?

Yes, it includes specific three-phase patterns (Pre-LLM, LLM Call, Post-LLM) designed to ensure RAG data is filtered by tenant and correctly attributed after the model response.

What is the core principle of this safety skill?

The core principle is that identifiers (IDs) flow AROUND the LLM, not THROUGH it. This prevents the model from hallucinating IDs, leaking data across tenants, or being manipulated via injection.

What parameters are considered 'forbidden' in prompts?

Sensitive system metadata such as user_id, tenant_id, session_id, trace_id, and any raw UUIDs are restricted from entering the prompt context.

LLM Safety Patterns

Name: LLM Safety Patterns
Author: yonatangross

byyonatangross

•

보안 및 테스팅

Implements production-grade security patterns for LLM integrations to prevent prompt injection, hallucinations, and sensitive data leakage.

The llm-safety-patterns skill provides a robust architectural framework for securing LLM-driven applications within Claude Code. It enforces the critical security principle that system identifiers should flow around the LLM rather than through it, effectively neutralizing risks associated with cross-tenant data leakage, ID hallucinations, and prompt injection. By providing standardized patterns for pre-LLM filtering, regex-based prompt auditing, and deterministic post-LLM attribution, this skill ensures that AI integrations remain safe, auditable, and grounded in multi-tenant environments.

주요 기능

0169 GitHub stars

02Multi-tenant isolation through pre-LLM RAG filtering

03Context separation patterns to keep sensitive IDs out of prompts

04Deterministic post-LLM attribution for accurate source tracking

05Comprehensive output validation for schema, toxicity, and grounding

06Automated prompt auditing for forbidden parameters and UUIDs

사용 사례

01Securing RAG pipelines against prompt injection and hallucinated references

02Building defense-in-depth architectures for production-ready AI services

03Preventing cross-tenant data leakage in multi-user AI applications

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add yonatangross/orchestkit llm-safety-patterns

For use in Claude.ai and ChatGPT

Download Skill