Microsoft research finds Microsoft AI products may never be secure
Microsoft CEO Satya Nadella is going all-in on AI. Earlier this week, he announced that the company’s developer division (which makes developer tools and compilers) has been folded into a new unit …
big week for renaming things at MS
User: merrily uses Office, words in Word, excels at Excel.
Microsoft: Hey user, buy 365 because we want all your data.
User: fuck off I like Office.
Microsoft: ok then
User: ugh FINE
Microsoft: Hey user, pls use Copilot we really splurged on that one.
User: bro I’m busy doing spreadsheets rn
Microsoft: ok then
User: oh ffs
I'm waiting for the Microsoft 365 Copilot Office Live Premium Pro Clippy Edition.
Dumbest shit I've seen this year.
*...but don't anyone worry about that! We're going to take control of a nuclear power plant and keep pushing them to you whether you want them or not! We've never let the sanctity of users' data stop us before, and we're not about to start now!"
You’re not wrong. MS announced adding CoPilot or whatever the fuck it is called this week to the Office 365 subscriptions for an additional fee, while hiding the subscriptions without under unclear cancellations and shit.
So, let's go over this shall we.
Humans were always the weakest link in the security chain. Why? Because humans aren't logical and can be tricked with words and ideas.
So we've developed this new type of computer program that "thinks" and speaks naturally like a human, right? Responding naturally to human conversation with it.
The issue is (as I've said before) that we've essentially created a computer program that is just as fallible as humans.
In other words, no shit simple prompt engineering works. There's no way to "secure" a human brain from dripping out things it shouldn't on accident, and by extension, there's no way to "secure" an LLM "brain" because they operate in a somewhat similar manner (or at least appear to). Prompt engineering is just social engineering for computers. We've created a computer that can be tricked with words and ideas, just like a human.
Humans were the weakest link in security and we just made computers as weak of a link as humans. Who really thought making computers as bad at everything as humans was a good idea?
The issue is (as I’ve said before) that we’ve essentially created a computer program that is just as fallible as humans.
Id say it is worse, as we have more physical presence. We can think it rains, look outside and realize somebody is spraying water on the windows and we were wrong. The LLM can only react to input, and after a correction will apologize, and then you have a high chance it will still talk about how it rains.
We can also actually count and actually understand things, and not just predict what the next most likely word is.
But yes, I don't get from a security perspective people include LLMs in things, also with the whole data flows back into the LLM thing for training a lot of the LLM providers are prob doing.
I was going comment something about LLM can predict the next best word thereby forming a sentence and not inherently be bad because it makes sentences like humans do. You touched on that already. However, if something like copilot can leak private information then it stands to reason it is being train on private data unlike what we've been told. I know I know wolf eats sheep.
lol
I guess I will continue not using them
