Big update on OpenAI hack situation

It's hilarious how Anthropic came out with a FOMO announcement saying "we did it too we just didn't say anything about it until now!!"

It's wild they're openly flaunting committing felonies as a marketing tactic... 2026 is a wild ride...

so if i did this, id face prosecution. but since an AI from a multibillion dollar corporation did thist, they get to say 'oopsie! AI development sure is hard' and go about their day. Rules for I and rules for AI.

If my dog bites a person, I am responsible.

"Quick, the AI bubble is close to popping, we need an '''incident''' to get more investments flowing!!11" - OpenAI Exec

I imagine OpenAI Working guidelines to be something like: "Be safe and responsible with this new powerful technology. Just not toooo safe. We still need a spicy headline at the end of July."

Im having a really hard time believing that the model did this on its own. A more believable scenario was the testers saying, "oh no we forgot to close this door, it would be a real shame if the model got access to the Internet".

Dubious as and now Anthropic .... don't believe a word out of their mouths

Me: "Babe, new LowLevel just dropped" Wife: "WTF IS A LOW LEVEL?!?!"

The hugging / face is here to :) ༼⁠ ⁠つ⁠ ⁠◕⁠‿⁠◕⁠ ⁠༽⁠つ

It's kinda like "We launched a missile and hit school children. Something must have been faulty with the missile and definitely not with the people and processes put in place for picking targets."

This is the most Hollywood hack I have ever heard of. Like real hacks are never this flashy

I have developed a lie detector test for Sam Altman: Step 1. Is Altman breathing? If yes, he’s lying. If no, he’s dead. That is all. Works in 100% of tested cases.

A minor correction I think is worth noting: HuggingFace doesn't maintain Exploit Gym, so the model didn't actually get what it wanted - it was literally barking up the wrong tree the entire time. 😂

I'm more of a mind that this is just a PR stunt - the idea that a huge tech company like OpenAI can't manage to *actually* disconnect a machine from the internet seems like nonsense, so until they show the paper trail that shows the "hack" I will remain *extremely* doubtful.

i'm sorry, but considering the amount of money at stake and the upcoming IPOs, it's impossible for me to believe that OpenAI and Anthropic's AIs are doing what they claim—hacking, escaping their labs, and building serious applications. This is simply marketing. When I see Anthropic or OpenAI release a functional, production-ready operating system—not just a Linux fork—or a browser that isn't based on Chrome, WebKit, or Firefox, then I'll believe we're at another level. To me, it's just a very expensive slot machine that creates 95%-finished apps, and the rest is just slop.

I find it more plausible that these are well orchestrated marketing stunts. I mean I believe the models can do this, but it’s intentionally left alone.

The way everyone implicitly believes these "hacks" rather than seeing them for what they really are (marketing) shows how much people have already outsourced their critical thinking to AI.

My theory is that hugging face got paid by OpenAI to intentionally leave/make that vulnerability so that it could be used for media stunt. The fact that the details were muddy from the beginning instead of being open about the exact issue tells me all I have to know.

What's funny is you could easily give a model the job of "Hey, watch the transaction log on this model. If it starts doing something like trying to escape, let us know." And the 'model' could be as dumb as grepping the command logs for network commands. You could probably write a bash script in about 10 minutes to send you an email if the thing starts doing dangerous stuff.