External Data Extraction Attacks Against Retrieval-Augmented Large Language Models

Summary

This paper formalizes external data extraction attacks (EDEAs) against retrieval-augmented LLMs (RA-LLMs), where sensitive or copyrighted knowledge-base data can be extracted verbatim. The authors propose a framework built from extraction instruction, jailbreak operator, and retrieval trigger, and implement an attack called Secret. Across 4 models, including 3 commercial LLMs, Secret outperforms prior attacks and succeeds against all 16 tested RAG instances, extracting 35% of the data from RAG powered by Claude 3.7 Sonnet where other attacks yield 0%.

Topics

• Retrieval-augmented generation

Related items

• MediumCVE-2026-107430: VideoWhisper Site Manager for ChatGPT, Claude, Codex missing authorization checkSame vendor · NVD/CVE Database
• MediumCyberXero Uses Claude Code in Campaign That Stole Over 628,000 Ukrainian RecordsSame vendor · The Hacker News
• InfoAI’s quiet safety gatekeepers are stepping into the spotlightSame vendor · CNBC Technology
• InfoAnthropic is cutting off its internal evaluations from the internetSame vendor · The Verge (AI)
• MediumARTEX AI, Claude agents used in cyberattacks on South Korean banksSame vendor · BleepingComputer