Many-shot jailbreaking

This report presents the "many-shot jailbreaking" technique that exploits large LLM context windows by embedding many faux human–assistant dialogues in a single prompt to induce harmful responses; it documents experimental results showing effectiveness scales with number of "shots" and model size, explains the connection to in-context learning, and describes mitigation approaches (prompt classification/modification and fine-tuning) and their trade-offs.