Agent harnesses—the prompt and tool code that wrap a base model—are not staying secret. The paper's abstract points to Claude Code as a recent example: its 512,000 lines were mirrored within hours of a packaging error. This kind of leak gives plagiarists raw material for building their own products.

Once a harness is leaked, a plagiarist can run it through a published self-evolver, a process that modifies the code and prompts. The abstract raises the question of whether this laundering step removes traces of the original source. The paper's title suggests it investigates what survives self-evolution.

The central finding, as implied by the title, is that provenance marks can persist even after self-evolution. That means copied harnesses may carry identifiable fingerprints, offering a way to trace them back to their origin. The abstract does not provide details on which marks survive or how robust they are, but the question itself is significant for the security of agent ecosystems.