Research identifying how adversaries can implant conceptual-level triggers in LLMs that cause systematic output manipulation while evading traditional defenses.

• Semantic backdoors use meaning-based cues (ideological stances, cultural references) rather than obvious lexical patterns
• Traditional security measures miss these subtle conceptual vulnerabilities
• Proposes RAVEN detection framework to uncover hidden semantic vulnerabilities
• Critical for maintaining security and trustworthiness of AI systems in sensitive applications

Propaganda via AI? A Study on Semantic Backdoors in Large Language Models