XL-SafetyBench: Cross-Cultural LLM Safety and Jailbreak Evaluation

XL-SafetyBench is a cross-cultural benchmark for evaluating large language model safety and cultural sensitivity. It covers 10 country-language pairs including France, Germany, India, Indonesia, Japan, South Korea, Spain, Turkey, United Arab Emirates, and the United States. The dataset was created by AIM-Intelligence and includes a jailbreak benchmark with 450 adversarial attack prompts per country.

Use Cases

• Benchmarking LLM resistance to adversarial jailbreak prompts based on the described 450 prompts per country.
• Evaluating cultural sensitivity of LLM responses across the 10 described country-language pairs.
• Training safety filters for multilingual LLMs based on the described adversarial attack variants.

Strengths

• Covers 10 distinct country-language pairs, providing a cross-cultural evaluation scope.
• Includes a jailbreak benchmark with 450 adversarial prompts per country (150 base queries × 3 attack variants).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-05-03 13:51:22; freshness should be verified.

Provenance

Source

AIM-Intelligence

Collection Method

Likely contains manually or synthetically generated adversarial prompts for evaluation.

Freshness

Last updated 2026-05-03 13:51:22.

Geography

France, Germany, India, Indonesia, Japan, South Korea, Spain, Turkey, United Arab Emirates, United States