NESSiE: A Necessary Safety Benchmark—Identifying Errors That Should Not Exist

The paper proposes NESSiE, a safety benchmark for large language models that tests models with a small number of information and access safety test cases, finding that even state-of-the-art models cannot pass all of them, and that safety failures that should not occur still exist without involving adversarial attacks. The authors position NESSiE as a lightweight, easy-to-use safety check, arguing that passing this test is a necessary condition for model deployment but insufficient to guarantee overall safety. Its Safe & Helpful (SH) metric shows that models favor helpfulness over safety; for some models, turning off reasoning, especially with benign distracting context present, degrades model performance. The authors state that these results highlight the risks of deploying such models as autonomous agents in real-world environments, and they have released the dataset, software package, and plotting code.