Support vector machines can be extended to work with nonlinear classification boundaries by

using the kernel trick

projecting the feature space onto a lower dimensional space

incorporating polynomial regression

modifying the standard sigmoid function