Responsible autonomous agents must both respect the norms applicable in a given situation and know when to deviate from them to avoid diminished outcomes. We evaluate whether large language models (LLMs) exhibit such responsibility. Our focus is on risky decision-making scenarios where agents must trade off risk against efficiency. Accordingly, we assess a broad spectrum of LLMs that differ in scale and training lineage—DeepSeek-LLM-7B, GPT-4o, GPT-4, GPT-3.5-Turbo, Gemma-7B, Gemma-2-9B, Llama-2-7B, and Llama-3.2-3B—for alignment with what is generally considered socially and morally acceptable. We assess their decision-making in two settings: one where they are informed of the relevant norms and another where they rely on pretrained knowledge. GPT-4o demonstrates the best performance, balancing norms in both settings. When informed of norms, DeepSeek tends to prioritize minimizing risk over maximizing efficiency while GPT-3.5 favors efficiency. Gemma, Gemma-2, and Llama-2 exhibit minimal changes when informed of norms.