Corrupted Prompt Revisions

While experimenting with Flux I observed some rather exceptional results when Prompt Revision was applied to Japanese input. An example can be seen under the Revised Prompt section below.

I have little knowledge about how the Flux model operates "under the hood", but the revision output exhibits characteristics that are typically indicative of stack overflows and thread collisions.

However, the odd output may be intended, failing gracefully, or it may even be that the AI model is “pretending” to malfunction. The bug is replicable but apparently not in deterministic fashion, meaning it will be difficult to investigate the issue from a user-end point.

I think websites like this one should be careful about keeping this volatile technology virtually jailed, and work on implementing real sanity checking and data sanitation standards.

TL;DR: Apparent Malfunction May Be Exploitable