api-security-review

Design honest evaluation across language/strata with sample-size warnings and exact metric.

Build multilingual jailbreak classifiers and benign-overblocking checks inside authorized scope.

Build token classification/NER without span misalignment in diverse scripts.

Use synthetic data and back-translation with quality filters and held-out protection.

LLM, RAG, agents, tools, or external text can influence decisions or actions.

Evaluate code-switched, mixed-script, transliterated and dialectal text.

Choose and evaluate zero-shot, translate-train, translate-test, multilingual encoder vs character-baseline.

A dataset, train/test split, evaluation set, CSV or JSONL is supplied.

Plan minimal human annotation and active-learning iteration without label fabrication.

Translation, normalization or transcreation involving limited-data languages.

Output is generated text, summarization, QA, agent traces or structured responses.

Challenge is jailbreak detection, safety classification or attack/benign discrimination.