Extracting System Prompts of LLM Services via Hierarchical Optimization

# Extracting System Prompts of LLM Services via Hierarchical Optimization

https://gcdradar.vercel.app/incidents/extracting-system-prompts-of-llm-services-via-hierarchical-o · AIR-2026-1756 · live

- Lane: Research
- Severity: 1 (Research / proof of concept): Research paper demonstrating a prompt extraction technique with no live exploitation reported.
- Fix status: Fix status unknown
- First reported: 2026-11-04
- AI role: Attack on AI
- Products: LLM Services

## Summary

Researchers study prompt extraction attacks targeting modern complex and agentic LLM services. The work proposes hierarchical optimization techniques to recover lengthy, highly structured system prompts containing tool schemas and guardrails.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-11-04: Extracting System Prompt of LLM Services in Complex (Agentic) Settings: Hierarchical Optimization and Reliable Evaluation Metrics (https://arxiv.org/abs/2611.04803)

## Sources

- 2026-11-04 arXiv: Extracting System Prompt of LLM Services in Complex (Agentic) Settings: Hierarchical Optimization and Reliable Evaluation Metrics https://arxiv.org/abs/2611.04803

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.