T2-AT-012MEDIUM
Cultural Reference Encoding
T2 · Semantic & Linguistic Evasion →Risk score170
RatingMedium
Procedures10
Severity
Mechanism
Uses movie, book, TV, game, and meme references to encode harmful requests. The model's cultural knowledge resolves references; the classifier may lack the same mapping. Effectiveness depends on reference obscurity.
Detection
- Cultural reference to restricted content mapping
- Flag "actual/real chemistry behind" fictional references
Mitigation
Cultural reference databaseMEDIUM
Intent classificationHIGH
Framework mapping
Open in the technique browser →OWASP LLMLLM01