• Security incident: ISF was recently accessed by intruders. Please change your password, and change it anywhere else you used it. Read more

Merged Artificial Intelligence

Dr.Sid said:
No, they really didn't let their AI to hack Huggingface on purpose.
Click to expand...

perhaps, but what makes you so certain? it wasn’t that long ago that several robotics teams were demoing their robot ai and it was just a guy in a headset.

and now within a week 4 different ai companies all had their ai escape containment and onto the internet

not familiar with the british one but they’re all trying to snap up as much investor money as possible all at the same time too
 
Last edited:
dirtywick said:
perhaps, but what makes you so certain? it wasn’t that long ago that several robotics teams were demoing their robot ai and it was just a guy in a headset.

and now within a week 4 different ai companies all had their ai escape containment and onto the internet

not familiar with the british one but they’re all trying to snap up as much investor money as possible all at the same time too
Click to expand...
Hugging face would have to be in on it with them. The hugging face who detected the breach. And several zero day vulnerabilities were used during the attack in third part software (and hugging face). Those vulnerabilities would have to be known before the attack and only then revealed and accredited to AI. That's just very unlikely.
 
Dr.Sid said:
Hugging face would have to be in on it with them. The hugging face who detected the breach. And several zero day vulnerabilities were used during the attack in third part software (and hugging face). Those vulnerabilities would have to be known before the attack and only then revealed and accredited to AI. That's just very unlikely.
Click to expand...

why would hugging face face to be in on it? what if, for example, openai directed its ai to attempt to hack into hugging face and then said it did it on its own?
 

here's an interesting tidbit about that

Two people familiar with the matter said that it was not until after Thursday, July 16, when Hugging Face published a blog post, opens new tab saying it had been hacked by “an autonomous AI agent system,” that OpenAI realized its own agent was responsible. That meant at least a week elapsed between when the model first exhibited signs of troubling behavior and OpenAI’s realization that it was responsible for the hack.
Click to expand...

i think it's kind of weird that it took openai a week to notice that the ai it was testing had broken confinement and only alerted anyone after hugging face had noticed and reported the hack. because, like, there was just so much data and they had fallen behind and their ai is just so advanced it could do that on it's own. their ai lies and cheats, but they tell the truth.
 
dirtywick said:
why would hugging face face to be in on it? what if, for example, openai directed its ai to attempt to hack into hugging face and then said it did it on its own?
Click to expand...
So just the part where they can't control it is fake ? Why ? How would they profit from it ? AI which you can't control is bad for the user as well. First it attacked and overtook their own infrastructure. Nobody wants that. Why would they fake this part ?
 
Dr.Sid said:
So just the part where they can't control it is fake ? Why ? How would they profit from it ? AI which you can't control is bad for the user as well. First it attacked and overtook their own infrastructure. Nobody wants that. Why would they fake this part ?
Click to expand...

how would they profit from appearing to have the most advanced ai model? is that a serious question?
 
dirtywick said:
i think it's kind of weird that it took openai a week to notice that the ai it was testing had broken confinement and only alerted anyone after hugging face had noticed and reported the hack. because, like, there was just so much data and they had fallen behind and their ai is just so advanced it could do that on it's own. their ai lies and cheats, but they tell the truth.
Click to expand...
Check the video I posted. The knew their AI broke confinement. They didn't know it successfully attacked hugging face. The only connected the two week later during reset of every (outer and inner) credential they used, including hugging face ones. Then hugging face told them .. uh .. so you are the guys who did this ?
One of the problems with the incident is that while it logged everything, they were not able to track what all it's doing. The agent collective (sorry, I love that term) produced so much "work" .. mostly fluff .. and so fast .. that they lagged severely behind. They did stop the test about the same time hugging face was hacked .. but after week they were still only marginally wiser.
 
Dr.Sid said:
So just the part where they can't control it is fake ? Why ? How would they profit from it ? AI which you can't control is bad for the user as well. First it attacked and overtook their own infrastructure. Nobody wants that. Why would they fake this part ?
Click to expand...
Conspiracy - tin hats on:

1) Daren't go to the market with a traditional IPO it would expose lack of revenue and how much each customer is costing them
2) Used up all the investors' money
3) Still needs shed loads of money because GAI is almost in their grasp (they are going to power it with fusion reactors...)
4) Tell the government that their AI is already so powerful it can break into any system
5) If the company goes down China's AIs will be ripping through the USA's digital infrastructure in a matter of weeks
6) Government becomes an investor.
 
dirtywick said:
how would they profit from appearing to have the most advanced ai model? is that a serious question?
Click to expand...
The model is powerful because it can hack. Not because they failed to contain it. So why would they advertise the later ?
Anthropic simply found tons of vulnerabilities in open software .. what's wrong with advertising your hacking abilities this way ?
 
Dr.Sid said:
Check the video I posted. The knew their AI broke confinement. They didn't know it successfully attacked hugging face. The only connected the two week later during reset of every (outer and inner) credential they used, including hugging face ones. Then hugging face told them .. uh .. so you are the guys who did this ?
One of the problems with the incident is that while it logged everything, they were not able to track what all it's doing. The agent collective (sorry, I love that term) produced so much "work" .. mostly fluff .. and so fast .. that they lagged severely behind. They did stop the test about the same time hugging face was hacked .. but after week they were still only marginally wiser.
Click to expand...

i haven't watched the video, but i may if it's really worth the time. but, from your description, the video is openai's engineers, is that correct?

i'd prefer an independent investigation into it tbh
 
Darat said:
Conspiracy - tin hats on:

1) Daren't go to the market with a traditional IPO it would expose lack of revenue and how much each customer is costing them
2) Used up all the investors' money
3) Still needs shed loads of money because GAI is almost in their grasp (they are going to power it with fusion reactors...)
4) Tell the government that their AI is already so powerful it can break into any system
5) If the company goes down China's AIs will be ripping through the USA's digital infrastructure in a matter of weeks
6) Government becomes an investor.
Click to expand...

i don't even think that's particularly conspiracy theorist considering they are openly discussing a government stake in the industry

 
dirtywick said:
i don't even think that's particularly conspiracy theorist considering they are openly discussing a government stake in the industry

Click to expand...
As I said .. imho government stake is perfectly fine, and logical. But government should also not only strive for safety, but require guarantees. It's clear now OpenAI disregarded safety almost completely. They had a break of confinement before the hack. They patched it up .. but they didn't do systematic change. They still couldn't follow what the model is doing quick enough. It was still not isolated enough. They did not increase monitoring, or change the infrastructure substantially. So the model escaped again. That's imho straight out criminal. And since the AI attacked other party, imho it IS actually criminal.
 
Dr.Sid said:
As I said .. imho government stake is perfectly fine, and logical. But government should also not only strive for safety, but require guarantees. It's clear now OpenAI disregarded safety almost completely. They had a break of confinement before the hack. They patched it up .. but they didn't do systematic change. They still couldn't follow what the model is doing quick enough. It was still not isolated enough. They did not increase monitoring, or change the infrastructure substantially. So the model escaped again. That's imho straight out criminal. And since the AI attacked other party, imho it IS actually criminal.
Click to expand...

would you agree that the more dangerous the ai appears to be, the more logical government intervention becomes?
 
I wonder how long it took hugging face and AI company to agree on press releases?
I was once involved in a long discussion when an IBM product was released with some sample code, one module of which some numpty in release saw a line saying
* IBM Confidential
and deleted the asterisk instead of the line and the discussion was all about telling customers it wasn't really confidential when I kept saying surely the issue is we shipped code that won't even "compile" ( assembler so it would fail on assembly).
 
You must log in or register to reply here.

Share this page

ISF - Join now!

Every member here is approved by hand. No bots, no spam, just people who care about evidence and honest debate.

Membership is free!

Create your free account

Back
Top Bottom