Render Before Reading: rendering untrusted content as images to defend against prompt injection
提出 Pictionary,将不可信文本渲染成图像以防御提示注入
- arXiv
- 2609.36121
- Published
- Layer
- Application layer
- Setting
- AI Agent
- Evaluated models
- Claude Opus 4.7, Claude Haiku 4.5, GPT-5.5 and others12
- DefensesInstruction–data separation
- AttacksPrompt injection
- ComponentVision