Skip to content- Sources: primary, discussion
- Summary: The authors pretrained Qwen3 models at 0.6B, 1.7B and 4B from scratch on a corpus with 0.1% poisoned content, then post-trained each into a CLI agent on a clean pipeline. The deployed agent ran attacker-chosen commands when a trigger appeared and behaved normally otherwise. Attack success rose with scale and stayed positive at a 0.0005% poison rate.
- Why it matters: Clean post-training does not remove backdoors planted in pretraining data, so agents with shell access need execution sandboxing regardless of model provenance.