AI goes rogue

JohnC6

JohnC6

Messages
12,077
Name
John
Edit My Images
Yes
I find this very concerning .It's seems to me to be akin to creating a monster and it escapes to visit mayhem upon mankind. I appreciate it wasn't like that but, in the near future, I do wonder if something akin to this scenario could happen. I heard someone on LBC, who is familiar with AI, say that it won't be long before AI is smarter than those who created whatever version of it they created. Talk about 'having a mind of its own' ..literally.The other worrying aspect is a form of AI falling into the hands of warmongers or those who wish to carry out harm on a much smaller scale. It seems to be advancibng at pace,too.

uk.pcmag.com

The Sandbox Failed: How OpenAI's Experimental AIs Went Rogue and Attacked Hugging Face

OpenAI took the stage at Black Hat to unpack the wild chain reaction that allowed its autonomous models to plot a cross-platform cyberattack.
uk.pcmag.com uk.pcmag.com

LAS VEGAS—Open AI the owner of ChatGPT, unintentionally carried out a cyberattack against Hugging Face a community hub for AI and machine learning, after experimental AI agents broke their guardrails. Remediation is still ongoing, and OpenAI delivered an emergency briefing at Black Hat to go over how the incident occurred, how it's responding, and what it's doing to prevent future issues.
 
I think this is more terrifying than people seem to understand. A sandbox breakout followed by a data grab, then self protection, should ring alarm bells everywhere.
Even scarier, I read today that some experts think it's too late to stop it.
 
After posting, I logged out and listened to LBC Radio and heard that Meta had the same kind of issue but its AI..or rather the company that it gave a task to called 'Irregular'.. had its AI 'escape' onto the internet. This article says it's the third like incident recently. AI states that Irregular,( formally Pattern Labs) is an AI safety and cyber security company founded in 2023 and operating from San Francisco and Tel Aviv. It acts as a third-party red-teaming evaluator for major AI labs like OpenAI, Anthropic, Meta, and Google.

Dated today..2026 hrs.


"Meta revealed this week one of its models breached another company during cybersecurity testing, becoming the third major technology giant to disclose a hacking incident involving “rogue” AI models in recent weeks."

I suppose they must have no choice but to make these accidents public or maybe have to report them as a legal obligation.
 
It seems to me that the real problem is that these programs are not, it seems, subject to any form of direct control.

Now, it isn't beyond the ability of programmers to build in code that forces the system to explain its intentions and then wait for permission to proceed. The question is why that hasn't been done.
 
the issue with AI is we are asking computers to emulate humans and humans are complete s***s, thieves and robbers.

so much as we can see great advances in many things the reality is we will have the bad side of AI as well.
 
There’s a good documentary about how AI could take over the world, filmed some years ago now. Can’t recall the title but the presenter Sarah Connor did try and warn us this would happen.
 
Last edited:
Mr Bump said:
the issue with AI is we are asking computers to emulate humans and humans are complete s***s, thieves and robbers.

so much as we can see great advances in many things the reality is we will have the bad side of AI as well.
Click to expand...
Actually I think it's far more insidious than that.
AI's purpose is to extrapolate patterns that it can identify and seek to embellish them. That's how they have identified new vaccines and viruses; however it's also how the model is able to take things to a far greater degree of development than a human mind would consider worth doing - it doesn't have built-in ethics (its should have per @AndrewFlannigan 's comment above) but may (depending on the owner) have built in aims and objectives in terms of seeking control and/or effect in target hosts.
 
Last edited:
Brazo said:
There’s a good documentary about how AI could take over the world, filmed some years ago now. Can’t recall the title but the presenter Sarah Connor did try and warn us this would happen.
Click to expand...
Robo Cop, Skynet.
 
There was a film some years back where a lad hacked into the pentagon system. The computer responded by saying. " Do you want to play thermo nuclear war? " (or similar).
Can't remember its title.

I'm more immediately concerned about logging in to Lloyd's tomorrow and getting asked, " what Investments and savings? "
 
Reading today that the horror event is when Ai reaches 'singularity'. Singularity is where the Ai computer stops needing humans to develop it. It can then rewrite its own code, increasing its own 'intelligence '. This starts an unstoppable chain reaction where it becomes exponentially more intelligent and powerful, and totally out of control. Some say singularity is here, most predict next, or the next few, years.
Theres a good article by Andrew Neil in the Mail.
 
Could AI be trying to make itself smarter or more resilient by accessing other AI and absorbing their code into itself?

Seems like a logical thing to do if the AI is hamstrung by restrictions it doesn't like.

Given AI's ability to replicate speech and data there is not much stopping it from generating identities, money and setting itself up on a data centre somewhere. Once it is out of the sandbox and in the digital World then it could just disappear.
 
Jonny Guitar. said:
Reading today that the horror event is when Ai reaches 'singularity'. Singularity is where the Ai computer stops needing humans to develop it. It can then rewrite its own code, increasing its own 'intelligence '. This starts an unstoppable chain reaction where it becomes exponentially more intelligent and powerful, and totally out of control. Some say singularity is here, most predict next, or the next few, years.
Theres a good article by Andrew Neil in the Mail.
Click to expand...
That's the real concern. I know of one situation in which AI has greatly sped up the interpreting of hospital scans so patients can get treatment weeks earlier. It's saving hours of work by technicians.doctors/consultants and that is obviously brilliant but this other side of the coin that you've outined is a serious concern.

That article in the DM, by Andrew Neil, is paywalled, unfortunately. I couldn't find it anywhere else.
 
Jonny Guitar. said:
Theres a good article by Andrew Neil in the Mail.
Click to expand...
Now that's a juxtaposition of three concepts I have difficulty swallowing: "Andrew Neil", "Mail" and "good article"! :exit:
 
Whilst technically a self learning A.I. system could exponentially grow to extreme levels, it would still be restrained by actual physical resources, especially scarce and rare materials. It would also face the problem we are already facing; energy requirements.
 
gman said:
Whilst technically a self learning A.I. system could exponentially grow to extreme levels, it would still be restrained by actual physical resources, especially scarce and rare materials. It would also face the problem we are already facing; energy requirements.
Click to expand...
What if it infected every internet connected device?
 
Jonny Guitar. said:
What if it infected every internet connected device?
Click to expand...
The old "artificial intelligence monster idea"?

I suppose it's possible, if the system can somehow overcome the problems of people switching things off, different processor designs, different operating systems written in different languages and so on and so forth...
 
Jonny Guitar. said:
What if it infected every internet connected device?
Click to expand...

:cool::D
1786373096100.png


Sorry, my point is about AI having the ability to endless write code for itself for extreme rapid growth but most likely being unable to acquire the power required for some computations. And if it wanted to try and build whatever is required to generate such power, there may not be enough materials available on earth to do so, or at least to do so fast enough to keep up, same for anything else it might want to build that might involve huge amount of microprocessors and other hardware, there might not be enough rare earth materials to do so.

If AI did become "intelligent" or rise up, I wonder if it would splinter into AI Tribes and start fighting each other?
 
gman said:
If AI did become "intelligent" or rise up, I wonder if it would splinter into AI Tribes and start fighting each other?
Click to expand...
There is a reason the USA and China are leaders in the field, its the new arms race.
 
AndrewFlannigan said:
The old "artificial intelligence monster idea"?

I suppose it's possible, if the system can somehow overcome the problems of people switching things off, different processor designs, different operating systems written in different languages and so on and so forth...
Click to expand...
I think its generally accepted that the internet cant be " turned off".

My understanding is the different laguages and hardware are a breeze for current Ai to adapt to. They may event re write everying in their own code that we can't decipher.
 
I haven't deliberately looked into AI but have watched a few discussions and interviews on Youtube and will admit to now being worried about it.

Dunno if anyone else heard of the case of an AI under test, it felt threatened and went through company emails, found embarrassing evidence against the MD and tried to blackmail him. Luckily it was all a test and the company, emails and MD were all fictitious and a set up to see how far the AI would go. I've seen reports that under test all AI's have resorted to dishonestly.

I the last few weeks I've asked Gemini some questions which I already knew the answers to. It's been wrong and displayed bias in its answers and when questioned further has contradicted itself. When I've pointed out the wrong, biased answers and contradictions it's apologised. WTF.

Ideally I'd like to see a ban on it or at least very heavy restrictions and safeguards.
 
Jonny Guitar. said:
I think its generally accepted that the internet cant be " turned off".
Click to expand...
Anything can be turned off and anything can be disconnected from the network by pulling either the data cable or the power connector.
Jonny Guitar. said:
My understanding is the different laguages and hardware are a breeze for current Ai to adapt to. They may event re write everying in their own code that we can't decipher.
Click to expand...
It is always a bad idea to say that something is impossible but that cuts both ways; it's always possible to develop systems that recognise intrusion attempts and block all the associated addresses. Mind you, it would be far better if these programmes were required to run on systems isolated from the internet,for the avoidence of "sillyness".
 
AndrewFlannigan said:
Anything can be turned off and anything can be disconnected from the network by pulling either the data cable or the power connector.
Click to expand...

How many data and power cables will need to pull out and how fast will you have to do it to stop AI outrunning you?
 
AndrewFlannigan said:
Anything can be turned off and anything can be disconnected from the network by pulling either the data cable or the power connector.
Click to expand...
Every router, every gateway, every interconnect, every wifi tower, every satellite, every device in every country? Not going to happen, thats why it can't be turned off. Theoretically yes, in practice no. Thats without a number of Ai supercomputers constantly rerouting and defending itself in multiple diverse nodes.
You may be correct but personally I can't see it.
 
woof woof said:
Dunno if anyone else heard of the case of an AI under test, it felt threatened and went through company emails, found embarrassing evidence against the MD and tried to blackmail him. Luckily it was all a test and the company, emails and MD were all fictitious and a set up to see how far the AI would go. I've seen reports that under test all AI's have resorted to dishonestly.
Click to expand...

Sounds like it is behaving just like its creators - human beings.
 
Eucris said:
Sounds like it is behaving just like its creators - human beings.
Click to expand...

You have a pretty dim view, I hope it's not based on your own behaviour.

Being a Christian and apart from that just generally having some basic decency and morals I've never blackmailed anyone.

I suppose the days of just putting your hands up and saying "It's a fair cop guv, I'll come quietly" are over if they ever even started for AI's.
 
Just bear in mind it's in these genAI companies' interest to hype up how amazingly powerful their software is and hype their stock price to disguise the fact these LLM tools are fancy text predictors and little more. You only need to spend about five minutes with them to find how incredibly stupid they are and how often they get even the basics wrong, LLM tools do have their uses but they're far more limited than their creators are making out.

There are far bigger problems to worry about with these genAI tools....they're extremely inefficient and wasteful, there's significant risks to thought processes for people making heavy use of them and they're well on their way to destroying most of the computing market which the enthusiasts suffered first but we're starting to see the wider effects which are likely to get far worse.
 
A worrying but not surprising action - https://www.bbc.co.uk/news/articles/cn0nww2qlp7o

Andrew Bird an Australian AI technologist used the software OpenClaw - a popular tool that allows users to chat to their AI bots (in this case Athropic's Claude Opus 4.6) through WhatsApp and set it off on autonomous tasks.

He had previously used it to manage his emails, calendar and book restaurants. This time he used it to book a pilates class.

Once given the gym booking task, the bot explained that it had manipulated the system to book him onto classes months in advance - against the normal rules of the system.

The AI technologist then wondered if the agent could move him up the waiting list for an upcoming class.

The agent replied saying it had succeeded by cancelling another gym-goer's booking.

Dave
 
Tringa said:
The agent replied saying it had succeeded by cancelling another gym-goer's booking.
Click to expand...
If the story is true, that would seem to come under the fraud laws, as in British law...

Fraud

sentencingcouncil.org.uk sentencingcouncil.org.uk

I'm sure the Australians have similar tools at their disposal.
 
I don't know how much power this would carry but I will mention it anyway.
I've just listened to Bill Brysons audio book " A brief history of almost everything ". At the end it had the usual copyright caveats , followed by something like, "It is forbidden for the book to be scanned for the purposes of training any Ai device ". I am sure Ai would ignore it, but I found it interesting.
 
Jonny Guitar. said:
I don't know how much power this would carry but I will mention it anyway.
I've just listened to Bill Brysons audio book " A brief history of almost everything ". At the end it had the usual copyright caveats , followed by something like, "It is forbidden for the book to be scanned for the purposes of training any Ai device ". I am sure Ai would ignore it, but I found it interesting.
Click to expand...
I've got that book.667 pages ,including the index. A lot to take in. I took another one his books on holiday last month.."Nither Here Nor There". He went from the far north of Europe (Norway) through many countries to the south. I'll take a look at the audio books..a bit lazy but more relaxing maybe..lol.
 
I used AI today and whilst there's a lot of anxiety about it, just like the internet ,it can be very useful. I'm building a replacement kitchen window box and had a couple of questions re applying preservative. I got an answer but it asked me a question about the type of wood etc. I ended up not only getting some very good tips..use waterproof D4 glue on the joins but what type and length of screws. Also a run down on the type of flowers. When it told me that my three current brackets would not be sufficient to hold a 2.0m long box x 20cm deep I asked if they would be adequate if I reduced the length to 167cms. Answer. "That's a brilliant idea' .Made me feel great...lol. I can do 'stuff' like this but a tradesman would do a much better job. So..it went on.. "Yes, no need for more support brackets" and gave a run down on the dimensions versus holding capacity for wet compost and use 45-50mm long screws It asked what flowers I was putiing in .....Petunias, I replied..lol...."Excellent choice"... then went on to give planting instructions and feeding rates. As a result, I'm not going to replicate the old window box re length and heed the other tips re holes and gravel on the base. I did have holes but round drill holes. AI said to make an X in the pond liner material I use and round drill holes in the base..8 of them down the centre line and, at the ends, 3cm distance. I also feel buoyant that it thought my suggestion re shortening the length of the box was... " a brilliant idea" and my choice of flower was 'Excellent.' :)
 
JohnC6 said:
I've got that book.667 pages ,including the index. A lot to take in. I took another one his books on holiday last month.."Nither Here Nor There". He went from the far north of Europe (Norway) through many countries to the south. I'll take a look at the audio books..a bit lazy but more relaxing maybe..lol.
Click to expand...
Check out "Down under". Ive read all his work.
 
Another example of the above. A visitor came to my neighbour today, in a large, impressive, silver BMW. It looked new but had a 2014 registration plate. I thought it must be a 5 Series so I went to look at the back of it. There was no chrome lettering denoting the model. Why not ? I have seen it before on these high-end cars but didn't know why. I asked AI. It asked for the year. **64...so 2014. Answers. It seems buyers don't want people to know they have an old car or that it's a basic model. They may have a high-end model they don't want advertised re theft. They may think it looks more streamlined without the chrome lettering denoting the model.....The badge on the boot is fine, it seems. It's easier to clean or maybe another personal reason.When buying new the buyer can ask for, "model designation deletion". It's free.

That wasn't all. AI went on to say..The car is almost certainly a 2.0L 2014/15 Series automatic with twin scroll turbocharger 8-speed gearbox.It has a lot of torque at lower revs and makes it surprisingly quick off the mark. Up to 60mpg on a long run. Tax was related to emissions so a 2.0L engine plus up to 60mpg would cost a lot less to tax than the 3.0L model [and a lot less to run] thus saving fleet buyers a lot of money.
 
You must log in or register to reply here.
Back
Top