agent-sandbox

Execute AI agents in a kernel-enforced sandbox with Landlock or Seatbelt isolation.

Updated Jul 18, 2026

Search Tags:#security #isolation #sandbox #agent-safety #landlock #nono

Ready to run

npx skills add https://github.com/arthrod/conejo-skills --skill agent-sandbox

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: agent-sandbox
Source: https://github.com/arthrod/conejo-skills/tree/main/skills/agent-sandbox
Command: npx skills add https://github.com/arthrod/conejo-skills --skill agent-sandbox

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This skill mitigates the risks of running untrusted AI-generated code by providing a kernel-level, capability-based sandbox that prevents unauthorized system access, data exfiltration, and destructive commands.

Core Features & Use Cases

Capability-Based Isolation: Uses Landlock or Seatbelt to restrict filesystem and network access at the syscall level.
Safety Guardrails: Automatically blocks destructive commands like rm or chmod and provides atomic rollback snapshots for safe experimentation.
Use Case: When testing an AI agent that needs to perform file operations, use this skill to ensure the agent is confined to a specific directory and cannot access sensitive system files or network endpoints.

Quick Start

Use the agent-sandbox skill to run the command python3 process.py with read-only access to the current directory and network access blocked.

Frequently Asked Questions about agent-sandbox

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

FAQPage Schema

How do I sandbox AI-generated code to prevent filesystem access and network connections?▼

You can sandbox AI-generated code using kernel-enforced, capability-based isolation that restricts filesystem and network access at the syscall level. This prevents unauthorized system access and data exfiltration by confining execution to specific directories.

What is Landlock syscall-level isolation for running untrusted AI agents?▼

Landlock syscall-level isolation is a kernel-enforced security mechanism that restricts filesystem and network access for untrusted AI agents. It provides capability-based sandboxing to prevent unauthorized system access, data exfiltration, and destructive commands during execution.

How do I block destructive commands like rm or chmod when testing AI agents?▼

To block destructive commands like rm or chmod when testing AI agents, use a kernel-enforced sandbox with built-in safety guardrails. These guardrails automatically prevent destructive system commands and provide atomic rollback snapshots for safe experimentation.

Can I execute Python scripts with read-only directory access and network blocked?▼

Yes, you can execute Python scripts with read-only directory access and network blocked by using a capability-based sandbox. It confines the agent to a specific directory, ensuring it cannot access sensitive system files or external network endpoints.

Does kernel-enforced agent isolation work with atomic rollback snapshots?▼

Yes, kernel-enforced agent isolation supports atomic rollback snapshots to ensure safe experimentation. This allows you to revert the system state after running untrusted AI-generated code, preventing permanent damage from unauthorized modifications.

What are the limitations of using Landlock for AI agent sandboxing?▼

Landlock for AI agent sandboxing relies on kernel-level syscall interception, requiring compatible system support. It is limited to restricting filesystem and network access, meaning it focuses on preventing data exfiltration and destructive commands rather than resource throttling.