Researchers have introduced CrackedPDFs, a new benchmark designed to test Large Language Model (LLM) systems against hidden prompt injection attacks embedded within PDF documents. This benchmark comprises over 29,000 generated PDFs, including both malicious and benign files, to evaluate detection methods. Initial evaluations show that a hybrid detection model, which considers document structure alongside text, achieves high accuracy, outperforming methods that rely solely on extracted text or structural information. AI
IMPACT This benchmark could lead to more robust defenses against sophisticated prompt injection attacks in document-processing AI systems.
RANK_REASON The item is a research paper detailing a new benchmark for evaluating LLM security. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- Connected Papers
- CORE Recommender
- CrackedPDFs
- DagsHub
- Gotit.pub
- Hugging Face
- Litmaps
- PromptGuard
- Pukaphol Thienpreecha
- ScienceCast
- scite Smart Citations
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
