LOKI: Alias-based Semantic Jailbreak Strategy for Large Language Models

초록

Existing filtering systems predominantly rely on surface-level, keyword-based defenses and are therefore vulnerable to evasion that exploits models' advanced semantic understanding. In this work, we propose LOKI (Lexical Obfuscation via Knowledge Insertion), a systematic strategy that exploits this vulnerability. LOKI introduces a novel semantic jailbreak technique, Semantic Alias Injection (SAI), which conceals malicious intent while preserving semantic coherence.In experiments, vanilla prompts exhibited an Attack Success Rate (ASR) of approximately 17-24%, and the automated baseline PAIR achieved limited performance ranging from 0.7% to 37.8% depending on the target model. By contrast, LOKI attained overwhelming evasion success with ASRs of 100.0% on Claude, 99.4% on Gemini, and 98.1% on OpenAI GPT-4o across the evaluated settings. These results demonstrate that current surface-based filtering approaches are inherently vulnerable to meaning-preserving attacks and indicate the need for new defense strategies targeting semantic-level intent detection for LLM safety. The implementation code is available at https://github.com/suh-hy/LOKI.Warning: This paper contains content related to harmful use cases. © 2026 Copyright held by the owner/author(s).

키워드