Davos discussion mulls how to keep AI agents from running wild AI agents arrived in Davos this week with the question of how to secure them - and prevent agents from becoming the ultimate insider threat - taking center stage during a panel discussion on cyber threats. "We have enough difficulty getting the humans trained to be effective at preventing cyberattacks. Now I've got to do it …
As I understand it, the problem with "AI agents" in their current guise is that to an LLM all inputs are prompts, i.e. they can't (because they simply don't) distinguish between data and instructions. I suspect the more guardrails you put in place to try to limit prompt attacks, the less flexible the system becomes.
Ultimately I think the only way to keep them from "running wild" is to simply not use them.
> i.e. they can't (because they simply don't) distinguish between data and instructions
Well, as statistical token predictors without logic, reasoning, programming, never mind intelligence, they simply guess wot a human the training data might do in a given context. So they indeed can't.
> Ultimately I think the only way to keep them from "running wild" is to simply not use them.
And not to build and invest the world's finite resources in them.
Too late for that though, sadly.
This post has been deleted by its author
I think we should focus on small purpose built AI/ML models. General ones are too sloppy. We don't need more than billion parameters in a model if we break them down and stick to unix philosophy. Companies already have kubernetes clusters full of microservices, add micromodels to that and a few gpus to accelerate the process if necessary and you are golden.
Some use cases of small models today: facial recognition, advanced ocr, voice recognition.
I don't know why hyperscalers care so much about big models if the end result is so consistent: slop.
Jessica, Hi,
AI agents arrived in Davos this week with the question of how to secure them - and prevent agents from becoming the ultimate insider threat - taking center stage during a panel discussion on cyber threats.
That opening sentence in the first paragraph is much better started and more accurately stated .......Proxy human agents of AI agents arrived in Davos this week with the question of how to secure them - ....
LLMs are not intelligent, they're statistical pattern matching systems that generate bullshit that may or may not be correct or true.
Here's an idea to make LOTS of money: invent a complicated but impressive bullshit generator, make $$$ from building and improving it, then make more $$$ from selling add-ons that make it more accurate, more correct, or less dangerous.
In general, the best solution to a problem is the simplest solution to the problem. The rush to use AI is just making everything more and more complicated and impossible to understand.
LLMs are not intelligent, they're statistical pattern matching systems that generate bullshit that may or may not be correct or true. .... Fonant
That surely then makes them virtually human, Fonant, .... and into programming humans with ideas to make lots of money for the complicated but impressive bullshit generators that they server with the sale of add-ons that make acceptance and resolution of the myriad problems causing their countless difficulties more accurate, more correct, or less dangerous?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
