Here's what just happened — in plain English.
Quick recap
Prompt injection is sneaky. Attackers feed the AI crafted text that hijacks what it does — no hacking required, just words.
It doesn't have to be visible. The malicious instruction can be hidden in a webpage, a document, or an image — stuff the AI reads, not you.
When it works, it's bad. Data leaks, unauthorized actions, users getting manipulated — real consequences from a text input.
It's not just hackers. Competitors, scammers, curious people — anyone with access can try it.
It's not jailbreaking. Jailbreaking bends the rules. Prompt injection hijacks the mission entirely.
RAG and fine-tuning don't fix it. Common belief, wrong answer. The vulnerability is in how LLMs process instructions — not in the data.