Because there was a discussion on reddit about this topic, I want to clarify that Johnson Lau explained how a check in the code prevents this attack.

So there is no real attack.

Also note that the subject of this thread has a question mark, which means that I'm asking the community for clarification, not asserting the existence of a vulnerability.

The segwit code is complex, and some key parts of the consensus code are spread over the source files (such as state.CorruptionPossible() relation to DoS banning, IsNull() check in witness program serialization, etc.).

Thanks again Johnson for your clarifications.

On Wed, Aug 24, 2016 at 10:49 PM, Johnson Lau <jl2012@xbt.hk> wrote:

Adding witness data to a non-segwit script is invalid by consensus:
https://github.com/bitcoin/bitcoin/blob/d612837814020ae832499d18e6ee5eb919a87907/src/script/interpreter.cpp#L1467

This PR will detect such violation early and ban the peer:
https://github.com/bitcoin/bitcoin/pull/8499

Another approach is to run the scripts of all incoming transactions. That's not too bad as you have already fetched the utxos which is a major part of validation.