The author of this analysis used ^ prompt and fed it to OSS LLMs. The question itself contains the answer. The prompt contains all the facts and insights the model needs to identify the issue. It removes the burden of finding those insights within the 4k+ lines of code in the original vulnerable file.

So the claim that OSS models can find those zero-day issues was a lie.