Definition
Seeding the huge dataset an AI model is trained on with malicious content so a hidden 'backdoor' is baked into how the model behaves. Researchers showed that corrupting just 0.1% of training data can implant a backdoor that later triggers remote code execution in a deployed agent and survives safety tuning. The poisoned behavior only fires under a specific trigger, so routine testing rarely sees it.
Why it matters
Any party that supplies or curates training data — including third-party data vendors — could plant a dormant backdoor no normal evaluation catches, making a model's trustworthiness unknowable.