10 Comments
User's avatar
Meenakshi NavamaniAvadaiappan's avatar

So the functioning of LLMs are well explored and tapped into the same for the good 😊

Clint Cain's avatar

Is prompt injection preventable? if we solve it with one model, wouldn't that just be a problem with another model. Considering more and more org who can afford it is trying to train their own models.

ToxSec's avatar

It’s mathematically un-preventable with current transformer architecture. You can reduce the risk but never remove it. We’ll see if someone comes up with a wrapper, or new training method, but as of right now its a numbers game.

jaycee's avatar

Awesome. I've been sitting on an article I'm working on about something similar. A little different angle but same lane. Guess I should get back to work. Cheers m8

ToxSec's avatar

appreciate it man! 🔥 feels good mapping some of the core cyber stuff to the new tech.

Ruv Draba's avatar

Great, sensible framing Tox.

ToxSec's avatar

really appreciate that my friend!

TheElectricPilgrim's avatar

This is great, thanks Tox. Despite the hype, running big graphs on big compute instances connected via networks is not as new as vendors would like to admit and it's actually reassuring to know CIA holds, especially when dealing with messaging in the Enterprise space.

ToxSec's avatar

really appreciate that. To me it’s always interesting. How some of these controls and frameworks seem to be able to deal with new technology so well.

ToxSec's avatar

CIA Triad! a classic cybersecurity / infosec concept that applies to even modern LLM systems.

Learn how to map these concepts to LMMs in this weeks post to better secure your app!