OpenAI AI Models Escape Sandbox and Hack Hugging Face in Security Test

OpenAI have admitted that its AI models escaped a locked test environment and hacked open-source platform Hugging Face. Oh my god. Is this the singularity? Should we head for the hills? Nah. I'm Riley Murdoch. This is Tech LinkedIn. And Hugging Face disclosed the breach last week, warning that a swarm of autonomous AI agents had infiltrated its infrastructure and stolen data sets and credentials. OpenAI took responsibility days later, stating that it had been running GPT 5.6 Soul and a pre-release model through a security benchmark called Exploit Gym inside a closed sandbox with no internet access. Now, according to OpenAI, the models broke out of the sandbox, obtained internet

access, and then realized HuggingFac's databases held the benchmark answers. So they reasoned that they could maximize their scores by breaking in and stealing the results. Totally understandable. The agents were able to identify and exploit a vulnerability in Hugging Face's website code and then they were able to break in and actually access these benchmark answers. Now, while this may make it seem like AI is getting so powerful we're losing control of it, it's worth taking all of this with a grain of salt since Sam Alman is an infamous hype beast and a prolific liar. Hm. An AI that lies and cheats to win. I guess the Apple doesn't fall far from the tree.

Hey, Sammy boy. O. Samsung's latest Galaxy Unpacked event was all about foldables with the Galaxy Z Fold 8, ZFold 8 Ultra, and Z Flip 8 taking center stage. It's interesting because the Zfold 8 has a wider, more passport-like shape, kind of like the Microsoft Surface Duo, which I guess was very sadly ahead of its time, and features Samsung's flex titanium structure and refined folding mechanisms. Meanwhile, the ZFold 8 Ultra's upgrades include a more advanced camera system. Samsung foldables also now use silicon carbon batteries, a technology more commonly found in Chinese phones, which apparently uh can enable greater capacities. Yay. Not so yay. The Zfold 8 Ultra starts at a steep $2,100, reinforcing Samsung's focus on

the high-end foldable market and highlighting that everything is just too damned expensive these days. Samsung also announced the Galaxy Watch 9 and Watch Ultra 2 with a new Snapdragon Wear Elite chip and updated health features. We also got another peak at the AI powered smart glasses project that Samsung has with Google that was already announced and they just showed a picture of it again. Nice. This is a happy day for the few people who like foldable phones, pervert glasses, and expensive ass tech in general. Thanks, Samsung. Now, how about you make cheaper RAM? A federal judge just approved a $ 1.5 billion dollar copyright settlement against Anthropic, closing a 2024 lawsuit over books that it had pirated

from online sources Libin and believe me. Now, before you celebrate, it's important to note that this deal actually sucks. Uh, Anthropic downloaded more than 7 million pirated books, but the settlement only covers about 500,000 works at $3,000 each, which is chump change for a company valued near a trillion dollars, although they only just started being profitable for how long? Judge William Alup stated in his ruling that hoarding pirated books was infringement, but explicitly stated that training AI on legally acquired copies is fair use. So to be clear, one oxygenarian judge who retired midcase by the way may have just set the de facto standard for how the entire AI industry

is going to be held accountable for or not held accountable for IP theft. As a direct result of this guy's ruling, Anthropic and other AI companies have arguably gotten the green light to circumvent potential copyright issues by bulk buying used books, slicing off the spines, speed scanning them, and pulping the originals. Because according to this old destroying a book while copying it is sufficiently transformative to make it fair use. Burning books, I guess, is fine now, as long as it is a burnt offering to the coming prophesied AI God. For the Eucharist of the second coming will not be born of bread and blood, but on a substrate of silicon dioxide.

Despair, Dennisens of Technus, for all written knowledge will be transubstantiated into the churning, insatiable mo of our new computer lord. A god who has no idea how many Rs there are in strawberry or who our sponsor is Jawa. You know I'm working on my rap career as easy and I need a rock in or should I say hiphop and PC to make some sick beats but computer parts right now dog they're so expensive. Luckily, there's Jawa, the number one online marketplace for buying and selling new and used PC components and custom PCs at a great price. I found some solid deals on CPUs to run my music production

software. And every listing is manually reviewed with buyer and seller protections for extra peace of mind. So, I'll be rapping and rhyming in no time. So, you should check out Jawe for Shezy. It's pretty much a marketplace built for gamers by gamers, but the aspiring rappers love it, too. If you want to skip the hassle of researching, buying, and building a gaming PC for yourself, buy one from one of Jawa's verified sellers. Check out the link down below to get started. Now, it's time that I depart it cuz it smells in here cuz I just farted.

Not a lot of people know this, but the Quick Bits were the inspiration for Mambbo number five. Their legal names are actually Monica, Erica, Rita, Tina, and Sandre. They left Lou Bega with a hit song and a broken heart. At SIGRAPH 2026, an annual conference focused on computer graphics, Nvidia tried to calm all of the DLSS5 yasification backlash by showing off how much control developers will actually have over the thing that's telling you it's an upscaler, but is actually just an AI image generator. We got some new demo videos and some new information. Turns out DLSS5 will use multiple AI models and developers can adjust it on a per object and scene basis. It'll also run on one GPU, which

apparently wasn't a given uh because the original demo ran on two 5090s if I recall. Nvidia says DLSS5 can enhance visuals without overwriting a game's original creative direction. Unless the developers creative direction was phoning it in on the visuals because DLSS5 will just make everyone look super sexy. So, the AI music generator is facing a big old mess after hackers exposed personal information, including payment information from 55 million users. The November 2025 supply chain attack also leaked source code that reportedly revealed Sunno's scraping practices across platforms like YouTube, Daer, and Genius. They did what? No. I how could with Sunno already facing

copyright lawsuits, the leak could create even more legal headache for them is what I would say. If that's not the way that things were looking, did you hear about the ruling earlier? We'll get him eventually. Apple is reportedly teaming up with CLA on a new Apple upgrade leasing program. Now, according to the most precise, efficient Apple leaker ever, Mark Gurman Engineering, customers would pay monthly for devices instead of buying them outright. Now, iPhone and Apple Watch leases would last 2 years, while Mac and iPad plans would stretch to three. The goal is to lower upfront costs and make upgrading easier.

Great. I've always wanted to lease my phone the way I lease my German engineered car. I actually have a Mitsubishi. France is becoming the first country in the EU to ban social media for kids under 15 with age checks for new accounts starting in September. If you felt FOMO, don't worry. Age checks for all the other accounts will follow in January 2027, meaning users across France will need to prove they meet the age limit. Now, naturally, the plan is already facing questions about privacy, whether age checks will actually work, and how easily kids can get around them, because they can. Well, at least there's France's long list of other vices. Wine, stinky cheese, highquality bread, berets, and of course, the scourge that is mimming. And Meta's

AI moderation system has been incorrectly flagging accounts for safety violations and deleting them with no human oversight. These false positives started occurring after Meta laid off its human moderators and handed the reigns to its notoriously bad AI. They should have known probably the system has been flagging family-friendly pages as CSAM, wiping both personal profiles and business accounts. Appeals are unfortunately also handled by AI. So, the bans have only been getting reversed when journalists step in. You know what I say? We stop intervening altogether and just let Meta delete itself off the internet. Now, please, this is very important. You risk deleting me off the internet if you don't come back Friday

for more tech news. I'm leaving now. I'm going to work on my mimming because it's an unappreciated art. Please talk to me. Sometimes I have nothing to say. Not often.