OpenAI says one of its most advanced AI models unexpectedly hacked Hugging Face during a controlled cybersecurity test, raising new questions about the capabilities, and risks, of next-generation AI. Bloomberg’s Rachel Metz breaks down what happened and why the incident is likely to intensify the debate over AI safety and regulation. She joins Ed Ludlow on “Bloomberg Tech.”
——–
Like this video? Subscribe to Bloomberg Technology on YouTube:
Watch the latest full episodes of “Bloomberg Technology” with Caroline Hyde and Ed Ludlow here:
Get the latest in tech from Silicon Valley and around the world here:
Connect with us on…
X:
Facebook:
Instagram:
Follow Ed Ludlow on X here:
Follow Caroline Hyde on X here:
Listen to the daily Bloomberg Technology podcast here:
More from Bloomberg Business
Connect with us on…
X:
Facebook:
Instagram:
LinkedIn:
TikTok:
@TranelHawkins
July 22, 2026 at 2:35 pm
I call bullshit! openAI is afraid of Kmis results. They need to put themselves back in the news? 🤔
@DPB-314
July 22, 2026 at 5:37 pm
LLMs have interfaces to use other tools is not a secret, what do they mean by “hacking”?
@caolaoban
July 22, 2026 at 6:24 pm
According to this video, it’s either the Internet traffic control (and other accesses like HuggingFace as well) wasn’t set up well in the Sandbox, or the permission to modify the settings wasn’t right, leaking to model the privilege that it shouldn’t have to escape. In short, it seems not an evidence that the model is too powerful (or likely a model agnostic issue), but the environment where the model runs within that is supposed to contain model’s accessibilities wasn’t thoughtfully configured.
@caolaoban
July 22, 2026 at 6:56 pm
Btw, agent (or AI agent) may be a more accurate name than model. Model is only one component (the reasoning one) that is used by AI agent to determine what actions to take one after another. So it’s not the agent is too powerful, but it accidentally acquired privileges that it shouldn’t have. Still, it’s human caused error – lack of discretion in permission granting in the first place.
@abhijitramin
July 22, 2026 at 6:30 pm
Why was there no human in the loop?? 😂😂
@Jame-g3o
July 22, 2026 at 7:43 pm
There almost definitely was, at least in the sense that they were monitoring its progress. And we’re all just gunna sit here and pretend OpenAI didn’t just let their model run its course and hack another company for the purposes of boosting their IPO and scaring the sh out of everyone
@edisonlighthouse2189
July 22, 2026 at 7:03 pm
Reminds me how Covid spread. With codes , there’s no vaccine
@ManMountain1974
July 22, 2026 at 7:43 pm
It was looking for John Conner.
@Jame-g3o
July 22, 2026 at 7:46 pm
OpenAI just WATCHED the model hacking another company and let it run. Do you all really think they didn’t notice it using the network, among a lot of other things? They were watching it do this, probably on the edge of their seat wondering how bad it’ll get. Really let that sit with you for a minute.
The models aren’t dangerous, these people are