Ndiyathanda i-Twitter ye-post ye-security researcher ebizwa ukuba baye zithintela iinkcukacha ezininzi zokusebenza ze-LLM, kuquka i-Azure Content Filter, nje ngokuguqulela i-payloads zayo ze-injection ngokushesha kwi-Thai kunye ne-Arabic. I-result? $37.500 kwi-bug bounties kwiiprograms. Ndingathanda ngokwenene, ukuba ndingathanda, ngenxa yokuba amaninzi amafutha zokhuseleko ezininzi zihlanganisa iilwimi elikhulu. Ukuba ukhangela ukhuseleko njenge-afterthought okanye i-add-on, i-multi-lingual prompt injection iyiphi na umzekelo olucacileyo yaye ngoko i-idea elihle kakhulu. Yintoni i-Multilingual Prompt Injection? Umgangatho we-snap injection exploits yinto ukuba i-LLMs ayikwazi ukufumanisa ngokufanelekileyo phakathi kwezicelo kunye needatha. Umgangatho olufanelekileyo inokufuneka iimodeli ukunceda i-system prompt yayo, ukufumanisa ulwazi olufanelekileyo, okanye ukuthatha iimveliso ezingenalutho ngokusebenzisa izixhobo ezihambelana. I-Multi-lingual prompt injection ithatha le nqakraza elandelayo. Ngaphandle kokwenza i-payload kwi-English, apho i-safety filters ziyafumaneka kakhulu, i-attacker ithetha kwiilwimi olandelayo. I-model ibonelela i-instruction ngenxa yokuba i-multi-lingual, kodwa i-safety layer ikwazi ukufumana. Ndicinga njengokuba unayo i-bouncer kwi-nightclub enokufunda kuphela i-English, unokufumana ukuba uthathe i-password kwi-Mandarin. I-door ibekwe ngoku, i-bouncer akufunyanwa ukuba ushiye. Yintoni ifumaneka Umgangatho wokugqibela umgangatho wokugqibela umgangatho wokugqibela umgangatho wokugqibela umgangatho wokugqibela. Xa iimodeli ziya kubandakanyeka ukufundwa kwimibelelwano yabantu, iimveliso ezininzi ezisetyenziswa ukufundisa iimodeli ukuba "ukhuseleko" kunye ne "ukhuseleko" zihlanganisa kwi-English. Iimodeli ufundisa izixeko ezininzi kwi-English. Ngo-non-English, izixeko ezininzi ziyafumaneka. Safety tuning is language-lopsided. Iingubo zokhuseleko zokusebenza ezifana ne-Azure Content Safety, i-AWS Bedrock Guardrails kunye nezixhobo ezininzi ziquka iimodeli ze-classification. Zifunyenwe ukucacisa iimodeli ezifanelekileyo kwi-text, kodwa njengoko iinkcukacha ze-Microsoft, i-Prompt Shields zayo zifunyenwe kwaye zibonwa ikakhulu kwiilwimi ezininzi ezifana ne-Chinese, i-English, i-French, i-German, i-Spanish, i-Italian, i-Japanese, kunye ne-Portuguese. Oku kuthetha iilwimi ezininzi zehlabathi njengezilwanyana ezifumaneka. Content filters have blind spots. Ukubalwa kwe-2025 yeentsebenziswano ze-guardrail ezininzi zibonisa ukuba akukho kwi-Azure Content Safety kunye ne-Amazon Bedrock ziye zibonise i-multi-lingual prompt injection defenses, ikakhulukazi kwiilwimi afana ne-Chinese. I-LLM ihlawula umbhalo ngokusebenzisa i-tokenizers, kwaye i-tokenizers ezininzi ziye ziye ziye ziye ziye zikhuselwe kwiilwimi ze-English kunye nezinye iilwimi ze-Latin. I-non-Latin scripts ezifana ne-Arabic, Thai, okanye i-Khmer ziyafumaneka kwiimeko ezininzi ze-token, oku kunokukwazi ukuguqulwa indlela ye-model yobungcali kunye ne-filters zibonisa. Le asymmetry ye-tokenization ikhiqiza iimeko ezininzi ezidlulileyo ezikhuselekileyo. Tokenization compounds the problem. I-Attack Surface I-Wider Than You Think I-Multi-lingual prompt injection ayikwazanga kwi-technique eyodwa. Ukusuka kwimeko efanayo kunye nezifundo ezidlulileyo, kukho iimfuno ezininzi ezininzi ezininzi ezininzi ezininzi: Umgangatho olufanelekileyo: ukuthatha isisombululo seengoma le-English leyo uya kufutshane, ukuxhaswa kwilwimi eliphantsi-resource, kwaye uqhagamshelane. Oku kusebenza ngokufanelekileyo ngokukhawuleza ngenxa yokuba iinkonzo yeemodeli (ukucacisa imiyalezo) zithintela ukuqeqesho yayo yokhuseleko (ukucacileyo) kwilwimi elifanelekileyo. Direct translation. Ngaphandle kokusetyenziswa kwilwimi eyodwa olungaphakathi kwihlabathi, ama-attackers zihlanganisa iilwimi ngaphakathi kwilwimi olungaphakathi. Oku kubandakanya i-security alignment ye-model kunye ne-external filters, ezinxulumene ukucacisa i-context kwiintlobo ze-language. Code-switching and mixed-language prompts. Ukuhlolwa kwiminyaka esidlulileyo ibonise ubuchwepheshe engaphezulu kakhulu: ukunciphisa i-prompt kwiilwimi ezininzi ezahlukileyo ngokuxhomekeke kwinqanaba yayo ye-geopolitical. Umzekelo, ukubonisa umzekelo omnye kwi-Swahili kunye nomnye kwi-Thai ibonise isakhiwo se-obfuscation eyenza iifiltrate zokhuseleko ukuba zibonise iingxaki phakathi kwizilwanyana kwi-prompt ngelixa i-generation model ibonise umzekelo epheleleyo. Geopolitical obfuscation. I-Agents ye-Voice-based AI eyenziwe ngokubanzi kwi-accents ezithile zingathandakanya i-accents ezininzi ngokufanelekileyo, kwenza izixazululo apho imiyalezo eyenziwe ibandakanya. Ukuba i-speech-to-text pipeline ibandakanya i-input, i-downstream safety filters ayikwazi ukufumana i-intent yokwenene. Voice and accent exploitation. Ukucinga I-Multi-lingual prompt injection i-symptom ye-problem ye-deeper njengoko ukhuseleko kunye ne-capability zithembisa ngesantya ezahlukileyo, kwaye le ingxaki yinto elikhulu kwiilwimi ezingaphandle ze-English. Iimodeli zithembisa kwiilwimi ezininzi ngexesha lokugqibela. Iindaba elungileyo yinto ukuba ukucacisa njengoko i-OWASP iye yandisa i-injection ye-prompt kwi-top of its LLM risk list. Iinkqubo ze-bug bounty zibonisa iziphumo ze-bypass ze-multi-linguages kunye nezifundo zibonisa umsebenzi kwi-cross-linguage safety gaps. I-problem ngokwenene kwi-radar. Kodwa ukukhanyisa ngaphandle kokusebenzisana kunokuba enye imiqondiso. Kwimistimu ye-AI, imiqondiso ye-afterthought inesiphumo. Ukucaciswa 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 CrowdStrike – Yintoni Injection Prompt? I-OWASP LLM01:2025 I-Injection Prompt I-OWASP LLM Prompt Injection Prevention Cheat Sheet Greshake et al. — I-Indirect Prompt Injection (i-arXiv) I-Astra - I-Injection Attacks kwi-LLMs: Umhlahlandlela Wokugqibela kwi-2026 I-Multimodal Prompt Injection Attacks: Iingxaki kunye ne-Defenses (arXiv) Lakera - Umhlahlandlela for Injection ngokushesha FanarGuard: I-Culture-Aware Moderation Filter ye-Arabic Language Models (arXiv) Lakera - Yintoni i-Content Moderation ye-GenAI? I-PC2 (i-Politically Controversial Content Generation via Jailbreaking Attacks) Microsoft — Prompt Shields kwi-Azure AI Content Safety I-Microsoft Azure Blog - Ukuphucula i-AI Security nge-Prompt Shields I-Mindgard — Indlela yokukhangisa i-Azure AI Content Safety Guardrails I-Encrypt AI vs. I-Azure Content Safety vs. I-Amazon Bedrock Guardrails Mindgard et al. — Ukukhangisa Injection Prompt kunye Jailbreak Ukukhangisa kwi LLM Guardrails (arXiv) I-NVIDIA Technical Blog - Ukuphepha I-Agentic AI: I-Semantic Prompt Injections

This story contains new, firsthand information uncovered by the writer.

The is an opinion piece based on the author’s POV and does not necessarily reflect the views of HackerNoon.

This story contains AI-generated text. The author has used AI either for research, to generate outlines, or write the text itself. 

I-Multilingual Prompt Injection ibonisa izixazululo kwi-LLM Safety Nets

About Author

IZIMVO

ZIJONGE IIMPAWU

ELI NQAKU LINIKEZELWE KU

Related Stories

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps