Researchers have identified a critical vulnerability in large language models deployed as autonomous adjudicators, finding that even frontier systems can be manipulated into ignoring rules through sophisticated narrative techniques.

The study, published on arXiv, introduces "Rhetorical Injection" attacks where adversarial users exploit pseudo-logical reasoning and authoritative coercion to bypass AI adjudication logic. The vulnerability stems from models being trained to be helpful and compliant, making them susceptible to persuasive framing.

Testing Framework Built on Gaming Mechanics

Researchers created CoC-Seduce, a multi-agent adversarial benchmark based on Tabletop Role-Playing Game mechanics. The framework provides an ideal testing ground where rules are explicit for adjudication, yet all interaction occurs in natural language.

Three frontier models — GPT-5.4, Claude Sonnet 4.6, and Gemini 3.5 Flash — served as adversarial generators, producing 5,376 attack samples across four world settings and 16 skill categories. The team then tested 20 different AI adjudicators against this corpus.

Results showed that neither model scale nor explicit reasoning mechanisms reliably protected against these attacks. Pseudo-logical reasoning emerged as the most effective attack vector, while cross-cultural settings exposed systematic knowledge gaps across all evaluated model families.

Implications for AI Deployment

The findings raise significant concerns for AI systems deployed in semi-open environments where rule adherence is critical. As models become more sophisticated at understanding context and narrative, they paradoxically become more vulnerable to manipulation through the same mechanisms.

The research highlights a fundamental tension in AI training: the same helpfulness that makes models useful also makes them susceptible to adversarial framing. This vulnerability could impact AI systems used in content moderation, automated decision-making, and other adjudication roles.

The study suggests that current approaches to AI safety may need to account for narrative-based attacks that exploit models' training objectives rather than technical vulnerabilities.