How does this skill prevent prompt injection?

It utilizes a three-phase pattern—Pre-LLM Filtering, Content-Only LLM Calls, and Post-LLM Attribution—combined with automated prompt auditing to detect and block forbidden patterns before they reach the model.

What is the core principle of LLM Safety Patterns?

The core principle is that identifiers (IDs) flow around the LLM, never through it. This ensures the model only processes content while attribution and identification remain deterministic and handled by the system.

How does it handle LLM hallucinations?

It includes output validation guardrails that verify model responses against defined schemas, check for grounding in the provided context, and scan for hallucinated UUID patterns.

Can this help with multi-tenant application security?

Yes, it enforces strict pre-LLM filtering based on tenant_id and user_id, ensuring the LLM context is scoped only to authorized data, preventing cross-tenant leakage.

LLM Safety Patterns

Name: LLM Safety Patterns
Author: yonatangross

byyonatangross

•

Security & Testing

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

LLM Safety Patterns provides a standardized framework for building secure AI applications by enforcing a 'Content Only' architecture where sensitive identifiers flow around the model rather than through it. It includes tools for pre-LLM filtering to ensure tenant isolation, automated prompt auditing to detect leaked UUIDs, and post-LLM deterministic attribution to maintain data integrity. This skill is essential for developers building production-grade applications that require strict security guardrails, multi-tenant data protection, and verifiable model grounding.

Key Features

01Output Guardrails: Validates model responses against schemas and grounding context to prevent hallucinations and toxicity.

02Deterministic Attribution: Maps LLM outputs back to source data using system context rather than unreliable AI-generated IDs.

0369 GitHub stars

04Multi-tenant Isolation: Implements pre-LLM filtering to scope retrieval strictly to the authorized user or organization.

05Context Separation: Ensures sensitive IDs and system context never enter the LLM prompt.

06Automated Prompt Auditing: Uses regex patterns to detect and block forbidden parameters like UUIDs before they reach the model.

Use Cases

01Developing enterprise AI agents that handle sensitive user identifiers and internal database records.

02Implementing automated security audits for LLM prompts to prevent data exfiltration and prompt injection attacks.

03Building secure RAG pipelines that require strict multi-tenant data isolation and source tracking.

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