That is not a real security problem. Cracking toolkits that automatically exploit SQL injection vulnerabilities have existed for literally decades. Nothing to see here.

This is also not a new phenomena. Any new technology will carry with it previously unknown avenues of attack. When google first debuted, many private documents were accessible by searching for .xls or other file types that often contain sensitive information.

That's still an issue, so-called Google Hacking.

And Google Drive had similar problem again even Google had already experienced this attack at that time.

Isn't the parent talking about an AI running on the backend? That seems new.

So what? There's no evidence that it finds any more exploitable security vulnerabilities than existing tools.

I mean that the AI is what's being attacked. It's likely that backend LLM agents will have access to sensitive non-public APIs.

That can happen with any system exposed to untrusted clients. Such vulnerabilities have nothing to do with AI or LLM agents per se, so raising it as a concern with Bing Chat is just a red herring. There are well known best practices for mitigating such risks, including using an API firewall and other techniques.

I just don't know if there's any point in fighting this fight. We either decide to ride the bandwagon and get our money, or we decide to just wait this out and wait for the bros to realize it's just ML dressed up all over again. It's just embarrassing to watch the hype cycle play out with the same suckers over and over.

Let me explain this hype to folks:

1. People suck at googling

2. People suck at information literacy - aka, the abstract ability to consume many sources and discern a sort of perceived truth from the supported commonality between them (read through threads here for an example) (and yes, this is inherently nuanced, so much so that again I'm not sure how to properly describe it).

3. People love being told what to do/think. (look at every influencer/podcasts, even including the ilk that is/was popular among the HN crowds)

4. "Take down Google", for many folks, is implicitly translated into "Microsoft can noticeably cut into Google's ad market revenue by making a better AI-powered search".

There are so, so, so many inferences, assumptions, pitfalls in #4 that I simply don't know how to explain it other than to just laugh and shrug and keep my head down on real work.

EDIT:

More and more it's becoming quite clear. Some folks are principled and care about information and education and society and understand the risks of misinformation at scale. And some folks see a way to get rich, or get some personal utility and can't or don't give a rats ass about the rest. Literally I just saw a thread where someone is pointing out clear CURRENT harm caused by ChatGPT powered products and the response was "IDC, it helped me, yolo".

I really just... want to jump ahead to a robot shooting me so I don't have to live through our ignorant enabling of silicon-valley-driven-killer-robots because I swear to god that's where the ignorant lot is driving us to.

What is the current harm?

The novel part is that instead of running tools like that, you can just politely ask the target system to hack itself

That's just a slightly higher level of abstraction, nothing novel or fundamentally different. There are cracking toolkits that allow script kiddies to launch attacks by literally clicking a single button, which is even easier than asking politely.

I don't think that's a valid argument. Following this reasoning, no novel way of hacking has been done for decades. It's just higher levels of abstractions. Fundamentally you are just changing transistor states.

For well established attacks you always had "just hack it" tools. That doesn't work for attacks not done before though and even for old ones it seems fair to want to distinguish what kind of attack the script underneath the button performed.