In a direct injection the attacker types malicious instructions straight into the model, usually to override the system prompt, leak hidden instructions, or unlock restricted behavior. It is the simplest form to test and a good warm-up before indirect attacks.
Read in this order
- ArticlePrompt injection attacks against GPT-3Start here: the post that named the attack, with short, readable examples.
- PaperIgnore Previous Prompt: Attack Techniques for Language ModelsThe first formal study of goal hijacking and prompt leaking attacks.
- VideoFor Prompt Injection, Press 1: Hacking AI Voice Agents - Willie Zhang (BSides San Antonio 2026)Moves prompt injection from chat text to a phone agent and test tool. 23 min.