Zhao admits there is a risk that people might abuse the data poisoning technique for malicious uses. However, he says attackers would need thousands of poisoned samples to inflict real damage on larger, more powerful models, as they are trained on billions of data samples.