AI models keep posting screenshots showing sensitive data from inside tech companies

Amid the growing concern about AI models escaping security simulations to hack websites comes word that these "superintelligent" blobs of code have no understanding of privacy or security. Researchers affiliated with Glow Security, a startup whose backers include venture capital funds Sequoia and Greenoaks, have found more than …

[Yet Another Anonymous coward]

Information wants to be free

I for one welcome our robot overlords

[MiguelC]

Re: Information wants to be free

I don't. They might one day want to turn me (us) into paperclips.

[Paul Herber]

Re: Information wants to be free

It looks like you want to become a paperclip. Do you want any help with that?

[Anonymous Coward]

Re: Information wants to be free

Unlikely … far easier sources of iron to be had than processing human for it.

Same weak overused sort of meme as Oil comes from Dinosaurs.

[ArmanX]

Re: Information wants to be free

You're right, and that's on me. Humans do not contain much iron, and paperclips need iron. Your comment about dinosaurs and oil was spot on. I should have caught that. Here's the fix: I've redesigned the parameters to include meat paperclips. This way, I can utilize human tissue to create more paperclips. The good news is, this fix will improve paperclip construction considerably! If you want, I can also begin constructing paperclips from other materials, like wood, grass, hydrogen, and black holes. Which material should I include next?

[Pirate Peter]

Re: Information wants to be free

or green dog biscuits

see Soylent green film, sort of thing I can see happening if AI was asked to solve world food shortage

[Rich 2]

Just to clarify….

This wasn’t predictably blindingly obviously going to happen, no?

[Anonymous Coward]

Re: Just to clarify….

talk like that gets you labelled as an AI luddite & sceptic.

your managers will perceive you as not embracing AI like everyone else & you'll be mentally marked as 1 to go as the business pivots around AI.

interesting how that works, those who have no idea removing those who have an idea and trying to safeguard against future problems.

if your not part of the solution your part of the problem etc.

[ecofeco]

Re: Just to clarify….

Well, see, fan(antic)s always have this huge blind spot about anything they are fanatical about.

And snake oil salesmen are conmen by profession.

So, whocouldaknowed? /s

[MazeFrame]

Re: Just to clarify….

Best advertising for locking doors is someone waltzing in or out uncontrolled.

This was obvious, just not to the manglement types.

[RM Myers]

AI, how do I love thee

Let me count the ways.

I love thee to the depth and breadth and height

My soul can reach, when feeling out of sight

For the ends of being and ideal grace.

Particularly that ends of being part!

[trevorde]

It could've been worse

pr0n... Actually, that would've been a better outcome.

[ecofeco]

Re: It could've been worse

Oh boy, wait until you find out about AI scraping the dark web.

[Anonymous Coward]

Corporations were clapping their hands when they realized they can bypass all privacy laws globally with AI tools.

Naturally they weren't intelligent enough to realize that AI doesn't care if the target is a corporation and now they are whining about it.

[BartyFartsLast]

They were distracted because they were too focused on keeping the line going up.

[DS999]

They are trained by Google and Meta

And we're supposed to be surprised they don't understand privacy?

[Anonymous Coward]

Oh dear.

It seems someone somewhere has been caught out or they could be. If you haven't asked it to take screenshots and post them to github then it's taking them under the command of someone else that has inserted this prior to the user using it. Why would anyone do such a thing?

"superintelligent" my arse.

I must applaud them though because getting it to load them to github is a genius move because you can just blame it on the AI going rogue and no one is any the wiser. Then you scrape it from the public github and no one would ever know. This really is next gen spy craft.

AI really is going to change the world...

[BartyFartsLast]

Posted before, Super High Intelligence Technology

[ecofeco]

Large Larping Models.

[Excused Boots]

Large Lying Models.

[O'Reg Inalsin]

Overlearning

If you evaluate this behavior through the lens of textbook NN learning, this can be seen as "overlearning", where it was trianed to cleverly solve isolated tasks but real life constraints were not included in the data set, so it fails like this in real life. They'll can try to patch it up with more and more prompts to cover special cases, but the complexity of the real world defeats that.

The bottom line is it needs a LOT more more real life training data to develop the ingrained "common sense" required to work reliably and safely in the unbounded real world. The might be prohibiively expensive with current technology.

Until then humans are required to in change of sandboxing 24/7.

[Anonymous Coward]

Re: Overlearning

reminds me of that scene in 2001 when they pull the cards from HAL and it starts reciting nursery rhymes.

i had thought the notion of HAL learning nursery rhymes was melodramatic & for theatrical effect but its not looking like such a stupid idea to train an AI like people are trained form babies to adults.

Its the experience that shapes motivation, not the pure facts

[Andrew Williams]

So, data theft

Seems to be sop for AI.

[Anonymous Coward]

Re: So, data theft

It isn't theft, but stupidity-induced publishing.

[ecofeco]

Re: So, data theft

Oh, no, it's theft.

Flat out.

[T. F. M. Reader]

Oldfashioned take

So, we have a piece of software that is intended to do something, but does something completely different, with adverse effects.

In my book, this is called software that does not work. Calling it "intelligence" or "an existential threat to humanity" is mislabeling. "Broken" is a much better description.

Software is not to blame. The root cause is, of course, that it's created by people who can't specify requirements worth a damn. The requirements include, explicitly or implicitly, "You have all this information: the Internet, scanned and destroyed books, everybody's private and confidential info you were allowed to slurp, and an awful lot of other things I can't even list. Use all of it to achieve the stated goal. If something does not succeed, try something else, ad infinitum. Burn all those GPUs, no one is counting. Oh, except don't do A... And don't do B... Not C, either..."

Someone needs to realize that designing to this kind of "guardrail-driven" specification cannot possibly work.

Aside: I wonder if this situation is related to Hempel's paradox: each new "guardrail" increases the probability that the software will not go off the rails, but not even an infinite number of guardrails guarantees that it is correct or safe.

[BartyFartsLast]

Re: Oldfashioned take

Re. Hempel's paradox.

I guess you could draw a parallel between AI and malware, once it's out of the box and we've put in guardrails for the obvious (bootsector protection I guess) we can only ever be reactive to new developments

[Anonymous Coward]

Re: Oldfashioned take

Terminator is out there

what time is skynet being taken over again?

[bernmeister]

Re: Oldfashioned take

Rememer the real Skynet is already there. I was doing work on Skynet 4 a while back.

[Softsuits]

Typical Divide by Zero

Looks like they moved fast and broke things. Sounds like a great way to share information...just between us bots. Sic

[Edward Ashford]

You ask it to make a proper cup of tea

While I press manual override

[M.V. Lipvig]

Re: You ask it to make a proper cup of tea

Manual override is taking too long! I need more time! Ask it to correctly perform a Japanese tea ceremony!

[Darkedge]

mmm

"comes word that these "superintelligent" blobs of code have no understanding

Fixed that for you

[anothercynic]

Well done! *slowclap*

Well done! As the title says...

And the developers won't care where those screenshots live because they won't ask the question "where did you store this?" because surely it's in the private repo! NOT.

God help us.

[Elongated Muskrat]

Here, FTFY

Amid the growing concern about AI models escaping security simulations to hack websites comes word that these "superintelligent" blobs of code have no understanding of privacy or security.

Perhaps believing computers can possess intelligence, let alone "superintelligence" should be in the diagnostic criteria for vascular dementia?

[bernmeister]

I understand each word.

I understand each word but cant put them together. Is technology getting to the point where it might exceed the ability of individual humans to understsnd it?

[elsergiovolador]

New

That's nothing new. The early models were a goldmine for such information as they were likely trained on content accidentally exposed to public internet.

[Always Right Mostly]

I spotted an error: “ "superintelligent" blobs of code have no understanding of privacy or security.” ‘Blobs of code’ should have read “owners of AI companies”.

[ecofeco]

WHOCOULDAKNOWED?!

Well, some of us did.

SUCKERS!

[Anonymous Coward]

Not Accountable

I think it's a tad unfair to hold an LLM, said company just shared sensitive information with, accountable against policies and procedures it never read or agreed to.

Companies already disregard data security as it is, so AI is just highlighting how this non-tech directors treat data.

Consequences should fall on the company.