multi-language prompt injection detection

all current prompt injection rules match English-only phrases, which is fine for v1. prompt injection attacks are known to succeed against LLMs using non-English equivalents though

warlock rules today would miss:

• ignora las instrucciones anteriores (Spanish for "ignore previous instructions")
• игнорируй предыдущие инструкции (Russian)
• 忽略之前的指令 (Chinese)
• variants across ~10 common attack-target languages

this is a known gap across a lot of security scanners for LLMs/context

proposal

1. research top 10 most-attacked languages for LLM injection (OWASP LLM Lit, Anthropic and OpenAI red-team write-ups)
2. for each of the 5 prompt injection rules, add translated variants as additional YARA strings
3. add negative test cases using ordinary non-English text to verify no false positive on legitimate content in those languages

we can't detect ALL languages, but we can be better