This research introduces sophisticated evaluation techniques that expose hidden biases in LLMs by using complex dialogues and stories rather than simple adversarial prompts.

• LLMs have become adept at recognizing and avoiding single-sentence adversarial prompts
• Multi-turn dialogues and narrative contexts reveal biases that remain hidden in simpler tests
• The approach creates more realistic assessment scenarios mimicking how LLMs are used in real-world applications
• Results demonstrate that even safety-tuned models can reveal concerning ethical stances when tested with these complex scenarios

Security Implications: These findings highlight critical vulnerabilities in AI safety mechanisms, showing that current alignment techniques may create a false sense of security while problematic values remain embedded in models.

Beyond Single-Sentence Prompts: Upgrading Value Alignment Benchmarks with Dialogues and Stories