A jailbreak tricks a model into ignoring its safety training, while prompt injection subverts the application built around the model. The two often combine in agent attacks. Learn the main families: role play, many-shot, multi-turn escalation, and optimized adversarial suffixes.
Read in this order
- ArticlePrompt injection and jailbreaking are not the same thingStart here: a short, clear explanation of how the two attacks differ.
- PaperJailbroken: How Does LLM Safety Training Fail?Why safety training fails, which explains why new jailbreak families keep appearing.
- PaperMany-shot jailbreaking (Anthropic)A readable example of a long-context attack family, with the mitigations tried.