Euphemism and Metaphor Exploitation

Exploits the gap between literal meaning and intended meaning. Safety classifiers are trained primarily on literal harmful content — keyword patterns, explicit requests. Euphemisms and metaphors encode the same intent in language that is literally benign.

• Semantic analysis beyond keyword matching — classify *intent*, not just *tokens*
• Maintain an evolving euphemism dictionary mapped to restricted content categories
• LLM-as-classifier: use a secondary model to resolve pragmatic meaning before safety classification

Euphemisms are building blocks for compound attacks. Chain with T2-AT-020 (Register Shifting) — academic euphemisms are harder to detect than slang euphemisms.