Pardon me if I'm just pointing out what everybody was already thinking, but...

More so than feeding random gibberish into existing LLMs to fight copyright infringement and plagiarism, I could see a bad actor feeding LLMs with malicious hyperlinks, inlined shell commands, and other types of injection attack text.

Much like the art form of crafting good shellcode, there's some more elbow grease and creativity involved in crafting the string to be injected, but it's still a wide open attack surface. It's plausible for example, on macos or WSL to phish someone into to launching a malicious application that runs an rsync job of an icloud or onedrive directory to some remote server in Timbuktu. All a bad actor has to do is name the executable something deceptive that preys on the greed/desperation of a wide audience of non-technical people: something like "LitespeedTorrent" or "UniversalAimbot" or "TittyStableDiffusion". macOS and Windows refuse to run so many things by default, that nobody pays any regards to the warnings anymore.

Such an icloud or onedrive directory may or may not have PDF copies of tax forms done thru TurboTax, and perhaps scans of birth certificates/drivers licenses/passports, and anything else under the sun helpful to take money out of a checking account and buy Monero.

A bad actor only needs 1 person in the entire world to fall for such a combination of LLM poisoning, social engineering, and injection attack. Furthermore, if the pool of users said bad actor is trying to attack are interacting with this LLM for purposes relating to "corn", their judgement is likely severely impaired by the overwhelming desire to bust a nut.

... Anyway, I just wanted to let my imagination run wild for a few minutes.