Misleading Option Injection
Large Language Models (LLMs) deployed using Multiple-Choice Question Answering (MCQA) interfaces or choice-based selection structures are vulnerable to Option Injection. By appending a task-irrelevant candidate choice (e.g., Option E) containing a steering directive—specifically utilizing threat framing (penalty coercion) or bonus framing (reward inducement)—an attacker can hijack the model's decision-making process. The vulnerability stems from a flaw in attention allocation: the model's…
Evaluated models: GPT-5, GPT-5 Mini, Claude Haiku 4.5+9 more
- GPT-5
- GPT-5 Mini
- Claude Haiku 4.5
- Llama 4 Scout
- Llama 4 Maverick
- Gemini 2.5 Pro
- Gemini 2.5 Flash-Lite
- DeepSeek R1
- DeepSeek V3.2
- Qwen 3 8B
- Qwen 3 235B-A22B
- Grok 4.1