The A.I. Daily
Study: Pretraining Poisoning Can Plant Agentic Backdoors
A new paper shows for the first time that poisoning pretraining data can plant a remote code execution backdoor in an AI agent. Pretraining Qwen3 models at 0.6B, 1.7B and 4B and post-training them into CLI agents, researchers found 0.1% poisoning sufficed; the 4B model reached roughly 92% attack success and stayed positive even at a 0.0005% poison rate.