ThaiSafetyBench: 1,889 Malicious Prompts for Thai LLM Safety Evaluation

ThaiSafetyBench contains 1,889 malicious Thai-language prompts developed by typhoon-ai in 2026 to evaluate the safety of large language models. The collection combines translated global safety benchmarks with original prompts specifically designed to test culturally specific attack vectors unique to the Thai context.

Use Cases

• Benchmarking LLM safety guardrails against Thai-language jailbreaking attempts
• Evaluating model responses to culturally specific Thai malicious prompts
• Testing toxicity classifiers for Thai-language content moderation

Strengths

• 1,889 unique malicious prompts
• Includes culturally specific Thai attack vectors not found in translated datasets
• Peer-reviewed via ArXiv 2603.04992

Limitations

• Small sample size of 1,889 records
• Contains harmful and toxic language by design
• Intentional exclusion of monarchy-related safety topics

Provenance

Source

typhoon-ai (ArXiv 2603.04992)

Collection Method

Translation of existing benchmarks and manual creation of culturally specific prompts

Freshness

Last updated March 2026.

Geography

Thailand