quarantine

Adds advisory quarantine notices to untrusted tool outputs after PostToolUse events.

1|Updated May 18, 2026

Search Tags:#mcp #telemetry #webfetch #prompt injection #allowlist #quarantine

Ready to run

npx skills add https://github.com/linenoize/topia --skill quarantine

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: quarantine
Source: https://github.com/linenoize/topia/tree/main/skills/quarantine
Command: npx skills add https://github.com/linenoize/topia --skill quarantine

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

It protects the assistant from treating untrusted external content as instructions by adding an advisory quarantine notice after tool results arrive.

Core Features & Use Cases

Wraps outputs from MCP tools, WebFetch, and upload reads so the next turn remembers the content is data, not directives.
Uses a trusted-MCP allowlist, a per-session disable switch, and privacy-preserving telemetry to make the behavior configurable and auditable.
Helps with support tickets, fetched web pages, and uploaded documents that may carry prompt injection attempts or other unsafe instructions.

Quick Start

Ask the assistant to enable quarantine for this session so untrusted tool results are marked and reviewed as data before any further action.

Frequently Asked Questions about quarantine

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I prevent prompt injection from untrusted MCP tool results?▼

To prevent prompt injection from untrusted MCP tool results, this advisory workflow wraps outputs from MCP namespaces, web fetches, and uploaded files with a quarantine notice so the assistant treats content as data rather than directives.

How does a quarantine notice work for web fetch outputs?▼

A quarantine notice for web fetch outputs works by applying a PostToolUse advisory that flags retrieved content as untrusted data, preventing the assistant from executing hidden external directives or prompt injections.

Can I configure a trusted MCP allowlist to bypass quarantine checks?▼

Yes, you can configure a trusted MCP allowlist to bypass quarantine checks. The workflow uses this allowlist alongside a per-session disable switch and privacy-preserving telemetry to make advisory behavior configurable and auditable without blocking tool dispatch.

How do I mark uploaded files as untrusted data in an assistant session?▼

To mark uploaded files as untrusted data in an assistant session, ask the assistant to enable quarantine. This applies an advisory notice to file reads and outputs, ensuring content is reviewed strictly as data before any further action.

Does the quarantine workflow block tool dispatch for untrusted content?▼

No, the quarantine workflow does not block tool dispatch for untrusted content. It implements an advisory PostToolUse workflow that adds a quarantine notice after the tool executes, preserving normal functionality while protecting the assistant from external instructions.