A security team is reviewing an LLM application's output validation approach. They currently use keyword filtering (blocking outputs containing specific words). What is the most important limitation of keyword filtering as an output safety control?

• A.Keyword filtering operates on lexical patterns and cannot detect semantically harmful content expressed without the blocked keywords - an LLM can explain how to do something harmful using synonyms, euphemisms, coded language, or step-by-step instructions that avoid trigger words entirely
• B.Keyword filtering is too aggressive and blocks all responses containing common words
• C.Keyword filtering is computationally expensive and creates unacceptable latency for production LLM applications
• D.Keyword filtering is effective but must be combined with rate limiting to be sufficiently protective