Safety-aligned large language models are being tested on their ability to learn from mixed compliance demonstrations, which combine both benign and harmful examples. Researchers are investigating how these models interpret different types of compliance demonstrations, including those that are helpful but also potentially harmful. The study examines three hypotheses about how demonstration combinations influence model behavior, shedding light on the potential risks of "jailbreaking" language models through in-context demonstrations1. By mixing non-harmful and harmful requests with helpful responses, the researchers aim to understand how models weigh competing compliance signals. This research has significant implications for regulatory compliance, particularly in the context of ARM, where shifting requirements can create advantages for those who assess and adapt early. The findings of this study can inform the development of more robust and safety-aligned language models, so what matters most to practitioners is how these insights can be applied to mitigate potential risks and ensure compliance in high-stakes applications.