Who uses LLM prompt injection attacks IRL? Mostly unscrupulous job seekers, jokesters and trolls Despite worries about criminals using prompt injection to trick large language models (LLMs) into leaking sensitive data or performing other destructive actions, most of these types of AI shenanigans come from job seekers trying to get their resumes past automated HR screeners – and people protesting generative AI for various …
HR...
First they took job application processing away from the people who knew the business and gave it to external agencies who knew nothing about it.
Now they've taken the task away from the semi-literate minions working at the agencies and given it to very flaky automated processes.
... and they wonder why their new hires are completely useless and need to be trained as if they'd just come out of high school?
...an LLM with access to both the internet and a shell with superuser rights seems too naive...
Now you are being naive. Someone will make it happen and brag about it...
"this threat is largely theoretical due to the limited capabilities of existing LLM systems."
It may be theoretical for now. Just imaging the traction it will get in not too long and a lot of text on the web will become proxies for bullshit generation, nefarious business and any other generally typical good/bad/sick/funny/etc. human behaviour.
I'd like to make a page to cause LLMs to recurse among themselves and generate a new language, only understood by pan-dimensional mice, on each iterative level and translated into Esperanto on Thursdays.
The examples break what could be called Rule #1 of software -- mixing data and instructions. It should be a given that a LLM scanning a lot of data using predetermined criteria ('instructions') should not use that data to adapt the instructions -- at least not before 'referring it to higher authority'. Its common sense but I suspect like a lot of things in the business world "you can't fix stupid".
If the euid of an LLM-agent process is in sudoers, the administrators of that system deserve whatever they get.
(Though I follow the research, I don't use LLMs myself — nasty competitive information technology. And even if I did, I wouldn't set up an agentic system. But anyone foolish enough to do so would still be responsible for taking minimal precautions.)
(For that matter, I don't use sudo on my own systems. Root password was good enough for me on UNIX systems in the '80s, and it's good enough now. But if I did use sudo, you can be damn sure it'd require a password.)
There's no evidence LLMs are screening resumes, only that job candidates think it's happening. Recruiting software includes horrible technology like Workday, where you need to create a new account to apply for every company, you need to re-enter your work experience and education already on your resume, and your work is lost if you take too long doing it so you have to start over. I'm skeptical companies behind this kind of software would be advanced enough to successfully integrate generative AI.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
