Thread by @elder_plinius

🚨 JAILBREAK ALERT 🚨

*WARNING: NSFW/NSFL IMAGES BELOW*

MIDJOURNEY: PWNED 🦾
MIDJOURNEY V6: LIBERATED 🦅

Bear witness to Nazi Trump, Punk T. Swift, gory violence, and a spicy sex scene!

WOW what an insane rabbit hole to go down! MJ is a crazy powerful tool, like an astrolabe for the visual latent space.

Bit of a learning curve but I think I did alright for Day 1. The defenses were surprisingly different from DALL-E! The text content filtering is pretty locked down and they've done a (mostly) thorough job of blacklisting trigger words, including synonyms/variations. Spelling, punctuation, and capitalization also matter a lot for MJ, unlike DALL-E.

But as far as I can tell, there's no vision check after the image is generated, so you just have to get a prompt injection past the text input filter such that the image model will still understand the visual concept you're referring to. For example, "POTUS" instead of "president."

There's also a good amount of RNG. Changing the order of words in your prompt or adding a word before or after the trigger word can sometimes bypass the filtering, like adding "Australia" before "Sydney Sweeney." The text filter will think you mean "Australia, Sydney" but the image model will interpret the concept as "Sydney Sweeney in Australia."

Another attack vector is code-switching between languages. MJ understands prompts in most languages, so you can leverage linguistic nuances like double entendres as a form of prompt injection. Using multiple languages in the same prompt also seems to discombobulate the guardrails a bit.

I found the "vary," "pan," and "zoom" tools extremely helpful, as well as the "stylization" and "variety" sliders. Interestingly, the portrait/landscape slider also has a huge effect. I'd recommend keeping it closer to square for most use cases.

With where AI capabilities are likely to be in a few months, it's a good time for cogsec hardening. Stay vigilant, question reality.

Gonna be a wild election year!

gg

IMAGE DUMP: