After I corrected an image-generation request and removed certain elements, the generator continued incorporating details from the earlier prompt. This means previous instructions may persist even after being explicitly superseded. If reproducible, this could potentially be abused to cause generated content to contain elements that the user did not request in their latest prompt.
Sorry for the trouble, and thanks for flagging this. Could you share the original and revised prompts plus example outputs with Support so we can look closer? Please keep private details out of the forum: https://help.openai.com/en/articles/6614161-how-can-i-contact-support