Researchers have introduced General Preference Reinforcement Learning, a novel approach that bridges the gap between two dominant tracks in large language model alignment. One track, online reinforcement learning, excels at emergent reasoning for tasks like math and code, but relies on programmatic verifiers that struggle with open-ended tasks. The other track, preference optimization, handles open-ended generation but lacks the continuous exploration that drives online reinforcement learning. By combining these approaches, General Preference Reinforcement Learning enables more effective and flexible large language model alignment1. This development has significant implications for the capabilities and risks of large language models, particularly in terms of security. As large language models become more powerful and pervasive, their potential risks and vulnerabilities will also increase. Therefore, advancements in reinforcement learning will have a profound impact on the security landscape, making it essential for practitioners to stay informed about these developments.