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

A new framework uses large language models to let researchers retrieve satellite and environmental data from cloud catalogues with natural language queries, according to a paper posted to arXiv. The system is designed to convert user intent into structured API calls, streamlining access to remote sensing archives. The architecture integrates three specialized agents: Guardrail, which enforces safety and policy rules; General-QA, which interprets user intent; and Recommender-Analyst, which generates schema-aware API calls [1][2]. The modular design allows the framework to be ported across different platforms by substituting API schemas, according to the authors [1][2]. They list environmental monitoring, disaster response, and climate analysis as target applications [1][2]. The paper describes preliminary experiments conducted under adversarial multi-turn settings. Prompt-level safety instructions improved robustness, but rare high-impact failures still occurred in API manipulation scenarios [1][2]. The authors note that these persistent vulnerabilities motivate the use of an intercept-level Guardrail agent and highlight the need for adaptive, system-level defenses that balance safety, usability, and cost efficiency [1][2]. Access to timely geospatial data has become central to tracking progress on international commitments. The United Nations Sustainable Development Goals, adopted in 2015 by all member states, include targets on climate action, clean water, and life on land [6]. A 2025 UN report found that only 35 percent of SDG targets were on track or making moderate progress, with nearly half moving too slowly and 18 percent in reverse [6]. Automated retrieval tools that lower the barrier to satellite imagery analysis could support monitoring efforts tied to those goals. The framework’s reliance on natural language interfaces reflects a broader push to make scientific datasets queryable without specialized programming. The paper’s authors argue that converting user intent into structured API calls establishes a scalable interface between human questions and geospatial infrastructure [1][2]. The preprint was submitted to arXiv on 13 June 2026 and has not yet been peer-reviewed [1].