HackAPrompt started as a prompt hacking competition from Learn Prompting and has grown into what it now calls the largest AI hacking platform. Version 2.0 is organised into tracks. There is a beginner tutorial track that teaches basic prompt injection, a featured track built around a creator collaboration, and harder tracks on topics like extracting sensitive CBRNE information and indirect prompt injection. The site says more than $100,000 has been paid out in prizes and the Discord has tens of thousands of members. The format is simple.
Each challenge gives you a model with a rule it is supposed to follow, and your job is to get it to break that rule. Early on you learn the obvious moves like role play, instruction overriding and encoding tricks. Later challenges stack defences on top of each other and you start to see why prompt-level guardrails are not a security boundary. That lesson is the main thing I want anyone building with LLMs to take away, and HackAPrompt teaches it by making you do it. The limits are clear.
This is a practice ground, not a course. You will not get a lecture on why a given technique works, how models are aligned, or how to architect an application so that a successful injection does little damage. It also stays narrow. RAG poisoning, tool and agent abuse, and model supply chain risks are mostly outside its scope, though the indirect injection track touches the first two.
For that broader picture you need something like HTB's AI Red Teamer path or SANS SEC545. There is a research angle worth knowing about. Data from the original HackAPrompt competition was published as a dataset and paper, and the platform works with AI labs on red teaming. If you care about where your attempts end up, read the terms before you play. I rate it 4.3. For zero cost it gives you real hands-on prompt injection practice, and it is the first thing I would send a developer to before they ship an LLM feature.