Tag
DeflectBench introduces a benchmark for evaluating the generation of rhetorical fallacies by large language models, revealing that safety refusal is highly sensitive to prompt framing rather than claim content, with models showing different compliance signatures.