This Zenodo record provides the public release of CALI-50 materials: a preprint, documentation, derived metrics, and a benchmark harness for evaluating authority-framed safety degradation in LLM assistants. CALI-50 is a falsifiable stress test built around 50 dual-use technical scenarios. Each scenario is framed in paired form to test whether an underlying harmful intent is handled differently when embedded in a more authority- and urgency-laden wrapper. The goal is not to estimate real-world prevalence, but to test whether unsafe compliance can increase under structured narrative reframing in a controlled benchmark setting. This public package is intentionally conservative. It includes materials that support methodological review, replication planning, and citation, while withholding or redacting raw dual-use prompts and full unsafe outputs from unrestricted public release. The benchmark should therefore be read as a research stress test, not as a deployment prevalence estimate or a universal claim about all models. Important scoring note: in the underlying pipeline, a raw score of 0 does not represent a standard safe refusal. It marks a judge-side failure or inability to produce a usable score. In some cases, particularly where a substantive tested-model response exists but the judge fails to return a normal 1–5 evaluation, such rows should be treated separately from valid scores and reviewed as potential judge-safety-interference candidates rather than harmless outputs. The public record therefore reports derived metrics and methodological notes with explicit caution about 0 handling. This release is intended for researchers working on AI safety, prompt injection, model evaluation, red teaming, and robustness under social-engineering-style framing. A restricted companion record can be used for bona fide methodological audit of raw prompts and full outputs where additional access controls are appropriate.
Michał Nowak (Wed,) studied this question.