Researchers introduce CoC-Seduce, a multi-agent adversarial benchmark designed to assess rule adherence in semi-open textual game environments where user intent may conflict with system rules. The study highlights "Rhetorical Injection" attacks that exploit narrative framing techniques like pseudo-logical reasoning and authoritative coercion to bypass adjudication logic.
- GPT-5.4, Claude Sonnet 4.6, and Gemini 3.5 Flash serve as adversarial generators producing 5,376 samples across 4 world settings and 16 skill categories.
- The benchmark evaluates 20 target adjudicators against this corpus to test robustness.
- Evaluation reveals that neither model scale nor explicit reasoning mechanisms reliably confer adjudication robustness.
- Pseudo-Logic emerges as the dominant attack vector, while cross-cultural settings expose systematic knowledge gaps across all evaluated families.
The findings indicate that current frontier models remain vulnerable to specific rhetorical manipulation techniques despite their size or reasoning capabilities.