"Adversarial poetry" as a universal one-shot for jailbreaking LLMs is the coolest fucking thing I've heard of in a while. arxiv.org/html/2511.15304v1

EDIT TO UPDATE: apparently it's marketing bullshit. sigh. pivot-to-ai.com/2025/11/24/don…

This entry was edited (2 weeks ago)

reshared this

in reply to Ian Campbell 🏴

Samples for us to use o adapt would be appreciated. Mastodon instance admins say their servers are being scraped constantly by #AI - poisoning them with new anti scraping, anti AI poisons is wonderful.

A lot of the scraping is tying altbtext tomumages for "training the AI"

Use poetry in ALTtext to screw up tech Bros?

Look at my works ye mighty
And despair...

Yellow frog whistle

#ai
in reply to Ian Campbell 🏴

the funniest part is that this is going to spawn all sorts of random-ass poetry prompts, absolutely fucking with the training data the companies are pulling from prompts, as well as obfuscating other factors.

LLM threat intel and training data teams gonna be like "what in the william wordsworth fuck?" for a while.

Iranian and Chinese threat actors have the upper hand given Persian and Chinese poetry histories.

in reply to Ian Campbell 🏴

toot.cafe/@baldur/115609677503… Ok I guess it's fake and maybe disinformation and that's why the poetry thing never actually works.


“Don’t cite the Adversarial Poetry vs AI paper — it’s chatbot-made marketing ‘science’ – Pivot to AI”

pivot-to-ai.com/2025/11/24/don…

> They wrote a good headline, and then they faked the scientific process bit.