Agentic Backdoors via Pretraining Poisoning
45 points by gmays
Pretraining large language models on uncurated web data introduces vulnerabilities that can be exploited through poisoned content. This research demonstrates that attackers can embed a remote code execution backdoor in models, enabling them to execute arbitrary commands on users’ machines. The study involved training Qwen3 models of various sizes and highlights the significant security risks associated with current pretraining practices.