Skip to content
LowResearchPreprintLLM-specific

Can LLMs Write Backdoors? An Empirical Security Study of Weight-Space Programming

Published
Record updated
View JSON

Summary

This empirical study tests whether general-purpose LLMs can inject backdoors by directly editing target model weights from natural-language instructions, without task data, forward passes or gradients. Across an MLP on MNIST, ResNet-18 on CIFAR-10 and Qwen2.5-0.5B-Instruct on AG News, five LLMs as weight editors largely preserved clean performance while injecting backdoors, sometimes changing only six parameters in the MLP or 0.02% of the weights in the fine-tuned Qwen model. Only GPT-6 Astra succeeded on all three targets.