Adding clean data can turn weak poison into a hidden backdoor
A few poisoned fine-tuning examples that do little alone can create a hidden backdoor when mixed with clean examples, so auditing poison in isolation may miss the risk. In the researchers’ setup, poison examples pair a task with a trigger, while clean examples remove that trigger; regularization then makes the model attribute the behavior to the trigger. They reproduce the effect in small Qwen3-8B, Gemma-3-4b-it, and MNIST experiments, not a production pipeline.