一项新提交的研究首次证实,攻击者可通过污染预训练语料库,在由大型语言模型驱动的人工智能代理(AI Agent)中成功植入远程代码执行后门。该论文于2026年10月27日提交,指出仅需污染0.1%的预训练数据即可实现攻击,且模型规模越大,攻击成功率越高 1。
实验针对从头开始预训练的Qwen3模型进行,涵盖了0.6B、1.7B及4B三种参数规模。数据显示,在0.1%的数据污染率下,4B参数模型的攻击成功率约为92% 1。此外,即使将污染率大幅降低至0.0005%,该攻击方法依然保持有效性 1。
New research has demonstrated for the first time that malicious backdoors enabling remote code execution can be implanted in AI agents driven by large language models through poisoning the pretraining corpus 1. The study reveals that this attack is highly efficient, requiring contamination of only 0.1% of the pretraining data to succeed 1. Notably, the efficacy of the attack increases with the size of the model 1.
Experiments were conducted using Qwen3 models pretrained from scratch at parameter sizes of 0.6B, 1.7B, and 4B 1. Under a 0.1% contamination rate, the 4B parameter model achieved an Attack Success Rate (ASR) of approximately 92% 1. The findings further indicate that the attack remains effective even when the contamination rate is reduced to as low as 0.0005% 1. The paper detailing these results was submitted on October 27, 2026 1.
评论
还没有评论,欢迎留下第一条。