Risk-Aware LLM Agents for Geospatial Data Retrieval: Design and Preliminary Adversarial Evaluation

A new LLM-driven framework for retrieving remote sensing data from cloud-based geospatial catalogues has been introduced, utilizing three integrated agents: Guardrail for safety enforcement, General-QA for intent interpretation, and Recommender-Analyst for API call generation. This modular architecture facilitates efficient interaction with satellite imagery and environmental datasets, supporting applications in environmental monitoring and disaster response. Preliminary adversarial evaluations indicate that while prompt-level safety instructions enhance robustness, there are still significant challenges in API manipulation that necessitate improved system-level defenses.