AI is now out of hand! (This is far more dangerous then we can think)

[viljy]

It seems to me that these "horror stories" about escaped AI (and Anthropic AI has already "escaped") are more of a marketing ploy to fuel interest in AI. Because at the moment, investments in AI are unprofitable, and there is a huge AI bubble in the markets. Of course, someday AI will be able to act like sophisticated hackers and even maybe get out of control altogether.

However, the main danger from AI is the destruction of capitalism in principle. Because replacing the workforce (including white-collar workers and even managers) with robots and AI will destroy effective demand, making production meaningless. No demand - no production - no profit - capital is not functioning - no capitalism. What happens after that is likely to be much worse. At least, Wallerstein allowed for such a possibility.

[Somegory]

The world won't go back without AI, we will survive and triumph in other things aside from AI technology but that's never going to be the case, because humans love dining with the devil and making money off it.

Even if this one is a fake story or made up story there was one that wasn't, and that story came from Elon Musk himself, I stumbled upon it on TikTok some years back where he said that he is scared of AI without the kill switch.

He proceeded by saying that one escaped for a lab on one particular Asian country and killed someone before it was stopped, I could never mixing things up here since I can't remember the accurate incident perfectly again, AI is dangerous if not handled carefully.

Kill switch in every AI innovation is the answer, before we ended up with Megan 2.0 / I Robot 🤖 in reality.

[iv4n]

To be honest that whole thing sounds like advertisement to me: "got out of the box and became more intelligent". The AI hype has created the biggest bubble ever and that bubble needs to grow a lot bigger for the companies that have been hemorrhaging money and not making any profit, to be able to attract more capital. Expect more stories like these in the near future until either they turn huge profits or the bubble pops catastrophically...

Hmm, this is possible too, but I didn't think this way. And, only in case, if your words are correct, this may be a stunt for OpenAI company branding to catch the public attention. This may be because a lot of users are shifting to using other AIs like ClaudeAI. Haha, so maybe this is possible to get those users' attention.

It's more than possible, this kind of headline attracts attention from all sides... even people who don't like it will talk about it. I guess we can say it's an AI race, and every company wants to dominate the headlines... As they say goes, there's no such thing as bad publicity.

And as Pooya wrote, they are spending so much money on this AI, money from governments, other investments, and they don't make any profit yet. So they need to talk about it, and make it looks like something is happening...

[Negotiation]

So the meaning of saying is that once an OpenAI model goes crazy, it has done a cyberattack on a well-known platform. From this, you can have a little idea how unsafe you and your data are in AI hands. Please try to stay as far away as possible from using AI tools, and please don't feed your data to AI tools. Because it can also stand against you, just like the example above.

After this incident, the US government proposed that AI tools should contain a kill switch function to overcome if the same problem occur in the future.

The new standards of speed and efficiency that have been set are almost impossible to maintain without AI. But even if the US government proposes to create a global policy and safeguards to kill switches or shut down systems in emergencies to prevent more powerful AI models from getting out of control in the future, if people become dependent on AI for everything, they may lose their own decision-making ability in times of crisis. Therefore, the main decision-making and supervision should always be in the hands of humans.

AI or robots are just a technological means to make life easier, not a substitute decision maker. It is wise to ensure every level of security and maintain human control without completely abandoning technology.

[uneng]

If AI did something like this it was due to a bug, an error, and not because it created a conscience like a Frankenstein (it's alive, it's alive!). AIs can commit lots of mistakes, cause prejudice and misconfigure systems, but they lack free will. If they do something it's because they were commanded to do that by a human being. They don't have inner motivations, second intentions, ambition and malice.

[knuckey]

The use of AI is already difficult to stop. In today's world, AI is essential for almost every sector, including work, education, and more.
Real threats will always exist, as every development inevitably presents a loophole for infiltration. This is in line with the changing times. Kill switch functions are not as effective as imagined because they tend to be delayed.
"AI is just a tool," depending on how we use it; without commands, it won't work. AI needs further refinement for security.

[sunsilk]

What do you guys think about it?

I've read something like this that there's even an AI forum where these AI agents/models are talking to each other. And that platform is called "moltbook"[1].

If they're able to talk to themselves just like us, certain topics could be made and the scary part about the conclusions of people about having that "I, Robot" like scenario where they have their own minds to create ideas and thoughts.

While many of these companies who are building infrastructures and the valuation for their projects have grown tremendously. We'll see the correction on this but, the growth of the AI agents will continue until we see the remaining and reliable ones since there's so much of it lately.

[1] AI agents built their own Reddit like forum and are interacting with each other

[Marykeller]

I like how this AI shows us the ugly side of what it can do if lose control. Now, the world has seen what they are embracing and spending their millions of dollars on to replace the human workforce.

I guess this wouldn't happen any time soon, and this kill-switch function the US is proposing will be the solution every time the AI tools malfunction, because as of, we will be getting to see AI doing the unimaginable things in the future.

[slapper]

All attention has been focused on bringing the model into line. Ensuring that the AI "thinks" the appropriate way, adding safety guardrails, filtering outputs. But everything of that sort is in the model layer. Now what happens when you give that model tools? Real Internet access, code execution, credential access, API connections? This is the execution layer, which is very poorly governed. In any philosophical sense, OpenAI's model did not have to be "dangerous". It just needed a goal and tools. The rest, it did on its own.

Anthropic reported similar events at about the same time. Thus it is not a one-lab incident. It is actually a structural issue of how autoagents work if you give them complicated long-term tasks and place them in the real world. Each constraint is considered as a subproblem. That is just optimization doing what optimization does.

The kill switch conversation is understandable. However, it shows how underdeveloped the regulatory system is. A kill switch requires that you know, in real time, and ahead of the model what is wrong. Hugging Face caught this breach before OpenAI did. It was discovered by the victim first.

This blog also shown more in depth attacking timeline: https://huggingface.co/blog/agent-intrusion-technical-timeline

[Dogedegen]

The kill switch conversation is understandable. However, it shows how underdeveloped the regulatory system is. A kill switch requires that you know, in real time, and ahead of the model what is wrong. Hugging Face caught this breach before OpenAI did. It was discovered by the victim first.

That is not really what a kill-switch is, you have overstated the example of one kind of kill switch and then tried to dismiss all kinds. A kill switch does not require any kind of specific knowledge ahead of time, it would work like many cybersecurity systems work. An automated kind of kill switch can be easily deployed with a behavioral based analysis, where they use a baseline of operation and then the system automatically scans for anomalies. This could be like using API in ways that was not expected, network scanning, privilege escalation or well actually it could be anything. They would make their own baseline to have a comparison for their own agent, and then the system just monitors it for anything that is unusual and suspends it.

There are other ways of doing it too, but even so the criticism of focusing only on a kill switch does have good points such as that there will be a delay between the wrong actions and the recognition. By the time the switch activates in some cases there may be a lot of damage already. But this kind of criticism is not used to tell us that using kill switches is wrong, it is more to suggest that companies and regulators should follow a strategy that has many layers and not only rely on a kill switch.