This is not possible. The limitation FME Workbench has is the following: It needs to know all Exposed attributes before runtime. This is also because dynamic attributes are not necessary to expose, because if you don't know that the attribute will be present there is no reason to use the attribute in one of the other Transformers.

If you need to keep the data with the dynamic attributes you can for example use a SchemaScanner in front of the FeatureWriter to write the Unexposed Attributes to your dataset.

If you need an attribute within a Transformer but are not sure your data will have the Attribute, you can still expose it, it will be exposed and set to <missing> if not present.