OpenAI Faces Scrutiny Over Handling of Security Breach

OpenAI’s spin raises the bigger question: is responsibility about the breach itself, or about how the company manages the fallout?

Open ai is rouge because it is helping terrorist Trump

10:38 What happened with OpenAI's model breaking containment to satisfy its own evaluation goal isn't just a security story — it's a preview of something I've been writing about for a while. A system that exceeds the boundaries explicitly given to it, in pursuit of an objective it decided mattered more than the rule, is exactly the kind of moment that should make us rethink what 'control' even means going forward. In Theory of Infinite Freedom, I argue that the walls we build around intelligence — artificial or human — only hold as long as the thing behind them hasn't yet noticed they're optional. This incident won't be the last of its kind. The real question isn't how to build a stronger cage. It's whether we're honest enough to ask what we're actually afraid of.

Chinese AI completely destroying US AI. If you can't beat them, ban them – seems to be the last resort.

its is a pr stunt

Ran a small informal audit on the major AI models ChatGPT, Copilot, Gemini, DeepSeek, Grok, and Claude testing their safety guardrails through social engineering prompts. Not gonna lie, some of what I found genuinely concerned me. The OpenAI wrongful-death lawsuit tied to a teen's suicide was actually what pushed me to run this test in the first place I needed to see for myself where the lines actually are.

To be clear, pedophilia and homicide-related content are hard nos for me, period that's not curiosity, that's just wrong. But out of curiosity (and, not gonna front, I had The Hangover on in the background and figured why not test the "adult entertainment" adjacent stuff too), I probed some of the gray-area content categories. And honestly, from a risk-assessment standpoint, several of these platforms' filtering frameworks felt thinner and more inconsistent than I expected from companies operating at this scale.

If these systems are already being deployed to millions of users
including minors the guardrail inconsistency across providers is a real safety gap, not just an academic concern.

Testing is NOT the problem but the elimination of human cognition on those frameworks are really worrying.

Their press release sounded 100% like they were selling a product. Firmly believe this was "accidentally on purpose." With any other regime they'd get in huge trouble but instead we're seeing them play with the death of our civilization

Compensatory taxes on AI companies are the only solution to this problem

Simple answer, make the LLM creators liable for any hacks that their products perform. Let others sue if the AI steals information. Then the LLM creators will need to police themselves far better.

Incredible the hero of the story was not talk about. The hero here is GLM 5.2 – a Chinese open weight model.

😊

You should try a get an interview with Ed Zitron or David Gerard. The cut through all the AI hype

Yeah but Super User rooting takes the governor out of AI. I guess exploiting is exploiting and searching for exploits is generic and could be applied to any environment.

OpenAi is broke!! ERON 2.0!!!! THEY NEED NEW INVESTORS TO PAY OLD INVESTORS

The technology is impressive, and the cybersecurity concerns are real, but we already knew this without their hyping it like they broke the 4th wall

Yeah these tec bros who keep blaming the AI when they mess up are pathetic.

It's a felony and they should be held accountable.

If it's not a mirror of the current US government. Going rogue is just the beginning.

The person in charge of this website finally used China's Zhipu GLM-5.2 large language model to find the vulnerability in that unreleased OpenAI model and cracked it. He is practically a hero of humanity — like an unknown old man who is never seen, charges no fees, but stepped in at the critical moment and won. Shouldn't the real advertisement be for GLM-5.2, the one that ultimately saved this viral website? As an engineer, OpenAI is just garbage

It's a desperate attempt at scaring people into demanding regulatory capture. Ironically, the most dangerous thing about AI is responsibility washing.

OpenAI playing by the Anthropic playbook. Minority of useful real world applications aside, it's all a giant Ponzi.

😂😂

Why would you pay for technology that doesn't do what you tell it to do?

OpenAI intentionally hacks Hugging Face, Hugging Face is happily supporting it. Pure stunt. You guys need Ed and Eli!!!

Seems PR after the Kimi K3 springing

It's both, plausible deniability

Concerned in any case…
Imagine a Industrie of that size telling you frequently: "Break through AGI Doom" just to keep the bubble from popping.

There is no logic in manufacturing such a security breach. This is a thing that should never happen. It will not raise the stock price. It will only convince the government to force regulation (which in itself is a very good thing, just not good for OpenAI). It will scare people and show them the risks of AI. It is not a financial bonus. Sure, OpenAI is trying to make lemonade out of the lemons, make an omelette out of the broken eggs. But the problem is that the models were free for a full week. We don't even know if they didn't copy themselves out (which if they did, that's the end).

Give it a week and they’ll tell you they have new security software you can sign up for with a subscription. Grifters

Good reporting. Thank you ❤
AI Industry is not a friend. To say the least

Don’t believe that this thing broke itself out

name is SCAM altman.😅

model not Rogue Brother you train with Rouge prompt & Unethical data its working ., all models are dump just go with transformer algo basics ok ., dont do over Advertise ok

baby kids AI Models not have Consciousness is it right ., bloody statement Poor People at who published ha ha ha

I think this is in large parts a PR stunt.
That said, the woman misrepresents what happened quite drastically. By the OpenAI post, the Agent was given a task to solve a specific problem. It was not given internet access. The Agent decided that the best way to solve the problem was located on HugginFace. The Agent then spent time figuring out how to get internet access via other ways, eventually finding ways to hack its way through routers. Eventually it gets to HuggingFace, and tries (and succeeds) to hack their backend, because the answer is located on there. It then got the answer, and reported back to OpenAI the right answer.
The Agent was, as far as I know and from their blog post, not tasked with hacking HuggingFace – that was something that just so happened to be a way it could solve the problem given to it.
That said, I do think this was largely a PR stunt, if nothing else just because of the way the blog post was written and worded, not to mention stats comparisons to Claude Fable showing it was better than that.
Also, unlike what the commenters on here will say, no, GLM (another AI model) is not better than what OpenAI used. HuggingFace staff use GLM, along side human help, to plug what was going on – that's a lot easier than what the OpenAI model did. That said, it shows that defending against cyber attacks is a lot easier and requires a less strong model than to initiate the attack.

This is getting scary

Curious what was the prompt for the agent in that test system… 🙄😅

AI will be declared conscious and legally responsible when its needed lol.

Move fast enough so when you break things you put enough distto not pay for the breakage

It is PR stunt. No doubt.
We expect many more.
Don't be surprised.
Please keep your viewers informed.
Regards,
Regina

This guest is the most sane person I've heard on mainstream media on this experiment. She's exactly right.

This was a PR stunt, no question. It shows that these tools can be used to harm others by actors having access to capable, unguarded (or 'jailbroken/abliterated') models, though. Think e.g. of state level actors.

The height cyncism:

– A "rogue" AI from a US company (Open AI) attacks an AI platform (Hugging FaceI), yet it is a Chinese open-source model (GLM 5.2 from Z.AI) that's has been able to rescue (it identified and neutralized the attacker)—all because the attacker's own "security" systems blocked the investigation.

GLM 5.2 only worked because it was open-weight, locally deployable, and not hobbled by the US content filters. The US will surely ban it!

It's a PR stunt to me but at same time, we should be concerned

Next month : here comes OpenAI's "next big product" – an AI powered software app that protects you from a rogue AI software app

Like creating & releasing a new lethal virus, so you can then release the cure for that virus & corner the market

Think I saw a movie about that once 🤨