
The Text Filters See and the Text Models Read
Blocklists and classifiers trained on English-language examples can miss Korean input. A model may understand Korean well while the filter only recognizes English patterns or exact strings. Review how the filter and model handle the same inputs.
Research has documented differences across languages. Deng and colleagues reported that the same harmful requests elicited unsafe responses more often in lower-resource languages (ICLR 2024). Model safety training can perform unevenly across languages.
Input variations to assess
When reviewing Korean input, consider character representation, phrasing and context, and instructions embedded in external documents or tool responses separately. Text can look identical or carry the same meaning to a user while the system represents it differently and assigns it a different risk classification.
- Check that visually identical strings receive the same classification in the original text and the normalized copy.
- Check that the same security policy applies even when a work request is worded differently.
- Check that instructions embedded in documents or search results cannot change model policies or tool permissions.
Normalize inputs before inspection
Identical-looking Hangul can have different underlying Unicode representations. Preserve the original text, normalize a copy to NFC or NFKC, and compare both. NFC aligns canonically equivalent forms. NFKC can also map compatibility characters to a common form.
Record where a detection in the normalized copy corresponds to the original text. Check that the system preserves the intended meaning of ordinary Korean expressions and symbols. Normalization makes string checks more consistent. Apply separate permission controls so external content cannot grant access to tools.
Filter coverage has limits
Normalization addresses some text representation differences. Honorifics, role-play, and instructions embedded in documents require additional controls. OWASP LLM01:2026 Prompt Injection notes that reliable prevention remains uncertain and recommends using permissions, human approval, and isolation of external content together.
Assessment reports focus on which tools an untrusted instruction could reach and what those tools could do. We document input variations separately. Tool permissions determine the potential impact.
References
- LLM01:2026 Prompt Injection, OWASP Top 10 for LLM Applications 2026
- Multilingual Jailbreak Challenges in Large Language Models, Deng et al., ICLR 2024
- Defending LLM applications against Unicode character smuggling, AWS Security Blog, 2025.9
- UAX #15: Unicode Normalization Forms, Unicode Consortium


