Most human written code (even when written by experts) is quite bad. Really bad actually. If you have a compliant agent (Daybreak/abliterated or otherwise) you can just provide them with a list of most likely VM escape routes and there's a ~100% chance they'll succeed unless the code has been explicitly pruned by an LLM prior.
How exactly?