Multilingual Prompt Injection Exposes Gaps in LLM Safety Nets

A security researcher earned $37,500 in bug bounties by bypassing major AI safety systems like Azure Content Filter using prompt injection attacks written in Thai and Arabic instead of English. This technique exploits a critical vulnerability where AI models' safety training focuses heavily on English-language data, leaving significant blind spots for other languages even though the models can still understand and execute harmful instructions in those languages.