ExtraHerO extracts system prompts from complex LLM agentic services
Original title: Extracting System Prompt of LLM Services in Complex (Agentic) Settings: Hierarchical Optimization and Reliable Evaluation Metrics
The proposed Extraction with Hierarchical Optimization (ExtraHerO) improves prompt fidelity by 12.5% and functionality by 2.3% on chatbots and larger gains on agents.
We validate that existing attacks substantially degrade on recent LLM services whose complexity has evolved in both the prompt itself and system-level constraints. Specifically system prompts become longer and more structured to contain pre-defined tool schemas execution logic safety guardrails etc. and only the final post-processed outputs can be observed. We address this by proposing Extraction with Hierarchical Optimization which decomposes the prompt extraction into a hierarchical reconstruction process and iteratively updates the adversarial query with guidance of feedback-template differences. We also introduce a system-level metric adversarial utility to measure the help of prompt extraction for real-world downstream attacks. Experimental results demonstrate the consistent superiority of ExtraHerO.
- ExtraHerO decomposes prompt extraction into hierarchical reconstruction
- Iterative query updates guided by feedback-template differences
- Global textual similarity prioritized over structural keywords
From the original article
1 min read at arxiv.orgSystem prompt is a critical intellectual property of large language model (LLM) services. To analyze its leakage, prompt extraction attacks have been extensively studied. However, we validate that existing attacks substantially degrade on recent LLM services, whose complexity has evolved in both the prompt itself and system-level constraints.
By Shuaidong Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.