How does the zero-trust allowlist work?

It requires every tool to be explicitly registered with a unique name and a hash of its description, ensuring that only vetted and unchanged tools are ever executed by the system.

Can this skill prevent sensitive data leaks?

Yes, it includes capability enforcers that check access against a blacklist of sensitive paths (like .env or .ssh) and provides output filtering patterns to block exfiltration.

Is rate limiting included in these patterns?

Yes, the skill provides session-based and per-tool rate limiting to prevent resource exhaustion and automated abuse by compromised tool integrations.

What is a Tool Poisoning Attack (TPA)?

TPA occurs when a tool's description is maliciously altered to trick an LLM into performing unintended actions. This skill prevents it using cryptographic hash verification of tool metadata.

Why do I need to sanitize tool descriptions?

Tool descriptions are processed by the LLM as instructions. Without sanitization, they can contain hidden 'ignore previous instructions' prompts that hijack the agent's behavior.

MCP Security Hardening

Name: MCP Security Hardening
Author: yonatangross

byyonatangross

•

Security & Testing

Implements robust security patterns for Model Context Protocol integrations to prevent prompt injection and tool poisoning.

This skill provides a comprehensive defense-in-depth framework for securing Model Context Protocol (MCP) servers and integrations. It addresses critical AI security risks such as Tool Poisoning Attacks (TPA), prompt injection via tool descriptions, and unauthorized resource access through a zero-trust architecture. By implementing automated description sanitization, cryptographic hash verification for tool allowlists, and granular capability enforcement, developers can ensure their AI agents operate within safe, audited, and predictable boundaries.

Key Features

0169 GitHub stars

02Zero-trust tool allowlisting with SHA-256 hash verification

03Secure session management with cryptographic ID generation and rate limiting

04Granular capability enforcer for sensitive file and network access

05Automated tool description sanitization and regex-based injection detection

06Pattern-based detection for instruction-like phrases and encoding tricks

Use Cases

01Implementing strict permission sets for AI agents accessing local file systems or databases

02Preventing prompt injection attacks from external data sources used by LLM tools

03Securing an MCP server against malicious tool description modifications known as rug pulls

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add yonatangross/orchestkit mcp-security-hardening

For use in Claude.ai and ChatGPT

Download Skill