Reinforcement learning from human feedback relies heavily on static reward models to align large language models with human preferences, but these models often lack the robustness to generalize to unseen domains due to the inherent diversity of human values. Researchers have proposed in-context reward adaptation as a solution to this problem, allowing for more flexible and dynamic preference modeling. This approach enables large language models to adapt to different contexts and preferences, addressing the limitations of existing multi-reward frameworks. The use of reinforcement learning in this context has significant implications for state-aligned activities, shifting the threat model from criminal to geopolitical1. This requires a different approach to security and mitigation, one that takes into account the complexities of geopolitical threats. So what matters to practitioners is that they must develop new strategies to address the emerging threats posed by reinforcement learning in a geopolitical context.