Jailbreaking attacks on large language models enable attackers to bypass built-in safety mechanisms, generating harmful or restricted outputs. Many-Shot Jailbreaking (MSJ) achieves this goal by leveraging demonstrations of multiple harmful behaviors. In this paper, we introduce Reward-Guided Many-Shot Jailbreaking (RG-MSJ), a novel approach that utilizes reward signals within the context to perform many-shot jailbreaking. Our method integrates question-answer pairs across multiple demonstrations with corresponding positive reward descriptions in a context window. For multi-turn conversations, we introduce negative reward descriptions for refusal behaviors, steering the model toward generating unsafe outputs in subsequent turns. Results demonstrate that RG-MSJ significantly outperforms baseline methods across single-turn, multi-turn, and function calling scenarios, performing well across various open-source models. Notably, with only 256 demonstrations, RG-MSJ achieves a 96.75% attack success rate on the Llama-3.1-70B model. Code is available at https://github.com/byerose/MSJRL .