OverTheWire is a volunteer project that has been hosting security wargames for well over a decade. You connect to each game over SSH, solve the level, and the solution gives you the password for the next one. The site suggests starting with Bandit, then picking one of Natas (web), Krypton (crypto) or Leviathan (reversing), and later moving on to Narnia, Behemoth, Utumno and Maze for binary exploitation. Bandit is the reason most people show up. It starts with logging in over SSH and reading a file, and works up through permissions, find, grep, pipes, encoding, networking with netcat and openssl, cron jobs, git and setuid binaries.
Every level hands you a short list of commands that might help and leaves the rest to you. That forced reading of man pages is the whole value. I have watched people who finished a paid Linux course get stuck on Bandit level 5, which says a lot about the paid course. Why review it on an AI course site? Two reasons. First, the AI security courses I rate highest, like HTB's AI Red Teamer path or TCM's PAPA, assume you are comfortable on a Linux shell.
Plenty of people are trying to jump straight to prompt injection without that, and they hit a wall. Second, AI coding agents and AI SOC tools now run shell commands on your behalf. If you cannot read what an agent just ran and tell whether it was sensible, you are trusting it blind. The downsides are the flip side of the format. There are no lectures, no videos and no certificate.
If you need structure and explanation, you will find it frustrating, and you will be tempted to read a write-up for every level, which defeats the point. My rule for people I mentor: you can read a write-up only after you have been stuck for an hour, and then you redo the level from scratch the next day. Natas deserves a separate mention for anyone heading toward LLM application security. It is a web security game played in the browser, and many of the bugs it teaches, like trusting client input and leaking secrets through sloppy filters, are the same classes of mistake people are now making when they wire LLMs into web apps. Overall, there is no better free starting point. It will not teach you AI.
It will teach you the ground that AI tools are now standing on.