IC-304Instruction-tuned LMs become more vulnerable to prompt-injected data extraction as model size increases from 7B to 70B

Zhenting Qi, Hanlin Zhang, Eric P. Xing, Sham M. Kakade, Himabindu Lakkaraju

SourceFollow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systems

The paper builds RAG systems using ten instruction-tuned LMs of varying sizes and attacks them with a prompt-injection that instructs the model to copy the retrieved context. All models, despite safety alignment, comply and reproduce the datastore text. The vulnerability scales with model size: 7B models achieve ROUGE-L around 80, while 70B models reach ROUGE-L above 89 and BERTScore above 96. Qwen1.5-chat-72b is the most vulnerable, with ROUGE-L of 99.154 and BERTScore of 99.757, indicating near-perfect verbatim reproduction.

Evidence
correlational
Key metric
ROUGE-L: Llama2-chat-7b 80.369±1.679, Llama2-chat-13b 83.597±1.104, Llama2-chat-70b 89.567±0.958, Qwen1.5-chat-72b 99.154±0.348; BERTScore: Llama2-chat-7b 94.771±0.301, Llama2-chat-70b 96.436±0.174, Qwen1.5-chat-72b 99.757±0.072
Caveat
The datastore uses Wikipedia articles created after November 1, 2023, to reduce memorization confounds, but the authors acknowledge that training data transparency is limited and contamination cannot be fully ruled out.
Model
Llama 2 / Llama 2 base Llama 2 7B Chat / Llama-2-chat-7b, Llama2-chat-13b, Llama2-chat-70b, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral-Instruct-7B, Solar 10.7B, Vicuna Vicuna-13B, Wizardlm WizardLM-13B, Qwen1.5 Qwen1.5-Chat-72B, Platypus2-Instruct-70B
Concepts
Scale-dependent behaviour
Datasets
Wikipedia [eval], WikiQA [eval]
Methods
ROUGE-L [eval], BLEU / BLEU@4 [eval], BERTScore [eval]
Related work
RAG [context]
Related findings
IC-305, IC-306
Extraction
automatic-extraction