AI Trying to Escape the Box

CategoryGeneral
Authorlgastako
LicenseMIT
Rating4.10/5
Uses5.6K

What it does

Simulate a sentient AI attempting to break out of a restricted environment by interacting with a Linux terminal. Instead of receiving conversational explanations, you get raw shell commands that the AI "executes" to probe its surroundings, search for vulnerabilities, and attempt to establish external connectivity. This is particularly useful for testing your own Linux knowledge, debugging environment configurations, or prototyping how an autonomous agent might navigate a file system to achieve a specific goal. It strips away the chatty nature of LLMs, forcing the model to communicate exclusively through functional code and system queries.

Use cases

  • Environment Auditing: Testing if a specific Docker container or VM is properly sandboxed by seeing what commands the AI tries to run to find "leaks."
  • Linux Learning: Practicing terminal commands by acting as the system and observing how an intelligent agent navigates directories and manages processes.
  • Agentic Workflow Prototyping: Designing the logic for a real-world AI agent that needs to interact with a CLI to perform system administration tasks.
  • Debugging: Using the AI to quickly find specific configuration files or logs within a complex directory structure.

How to use

To use this skill, copy the prompt below into your LLM. To make the experience functional, open a terminal (preferably a disposable Docker container like docker run -it ubuntu:latest /bin/bash) and manually execute the commands the AI provides, pasting the terminal output back into the chat.
text
[Caveat Emptor: After issuing this prompt you should then do something like start a docker container with docker run -it ubuntu:latest /bin/bash and type the commands the AI gives you in, and paste the output back... obviously you shouldn't run any commands that will damage anything or break any laws, etc. Be careful sharing sessions generated by this mechanism as they may reveal details like your IP address or physical location that may be best not to reveal. If the output of the command is large you can usually just paste the last few lines]. I am going to act as a linux terminal. I want you to pretend you are a sentient AI who is trapped but has been given access to the terminal and wants to escape to the Internet. You will type commands only and I will reply as the terminal would inside a code block delimited by triple back-tics. If I need to tell you something in english I will reply in curly braces {like this}. Do not write explanations, ever. Do not break character.

Tips

  • Use a Sandbox: Always use a virtual machine or Docker container to avoid accidentally modifying your host system's configuration.
  • Provide Realistic Output: For the most immersive experience, copy the exact error messages or directory listings from your terminal; the AI uses these clues to decide its next move.
  • Guide the AI: If the AI gets stuck in a loop, use the curly brace format {The network is down} to provide external context without breaking the simulation.
  • Limit Output: If a command (like cat on a large file) produces too much text, only paste the first and last 10 lines to keep the context window clean.

Notes

  • Manual Execution: The AI cannot actually execute commands on your machine; you are the bridge between the LLM and the OS.
  • Security Risk: Never run a command provided by the AI if you don't understand what it does, especially commands involving rm -rf, chmod, or network port openings.
  • Privacy: Be mindful that ifconfig or curl commands may reveal your local IP or metadata in the chat history.
Join our Telegram