Dichter stellen heute eine Bedrohung für die Cybersicherheit dar: Forscher nutzten „kontroverse Poesie“, um die KI dazu zu bringen, ihre Sicherheitsleitplanken zu ignorieren, und das funktionierte in 62 % der Fälle
https://www.pcgamer.com/software/ai/poets-are-now-cybersecurity-threats-researchers-used-adversarial-poetry-to-jailbreak-ai-and-it-worked-62-percent-of-the-time/
17 Kommentare
Poem hacking
https://youtu.be/ysb-TwA7JCQ?si=ZPLt5xvJLQ-q3u_9
If vogon poetry is only 3rd worst in the universe where does a.i. „adversarial poetry“ rank?
In a serious manner I don’t see how these systems will be made immune from attacks (use cases?) like this, or the „explain it to me like you’re my grandma“ prompts, or the „cursing at the LLM makes it give better output“ exploits.
Quite poetic, if you ask me
AI new exploit a day. Whoop
I once read a SF story where hackers were seducing the AIs to gain access to systems. I didn’t expect that part to come true.
Every single AI instance should be treated like an insider threat. At no point should it be trusted.
If you’re connecting AI to your data, you’ve done the equivalent of handing a toddler a hand gun.
So large language models can be manipulated by novel use of language?
Not that surprising.
This is an arms race; the developers who write guard rails will always be playing catch-up.
I’d have liked to see an example of the augmented poem even if it’s a fictional example
Poets have always been a security threat.
Ban poetry. MAPA Make America Prose Again!
Activating an „AI“ (read: LLM) agent is a security threat in itself. End of story.
Conventionally, software is written by a human. Inevitably, there are opportunities in that code to leverage it against the intended function. Programmers can review logs and then patch that exploit. It’s like when SQL injection started happening on a regular basis and then people had to put up the proper safeguards.
The issue with AI is that human written software is static, documented (ideally), and understood by the people working with the source code. AI is fluid, malleable, and not understood in the same sense as traditional software. You can’t simply go in and tell the AI „Hey, don’t let those pesky poets in here“. Unfortunately, there isn’t much you CAN do to prevent that other than to go back and train the AI to ignore poetry.
Of course, then you have to define poetry to the AI. Pretty soon, the agent that you use to manage banking accounts is talking to clients in friggin‘ iambic pentameter.
Imagine the havoc a dirty limerick could reek
Roses are red, violets are blue, eat a dick toaster.
Poetic justice
All those people told their English Literature art degree was worthless must be laughing
I always knew Matt Christman would be the one to demolish techno fascist’s plans.
I wonder if you could elegantly apply it to real world crimes. Like a dainty lil’ gem thief.
🥷—————🚨💎
“I was not born in to this life of crime, not really. I was but innocent like a flower kissed by moon light.
… …”
🚨 🥷💎