Post

Conversation

Matt Shumer
@mattshumer_
This is crazy... Read this blog from HuggingFace, written BEFORE they knew it was an OpenAI model that attacked them: huggingface.co/blog/security-
Quote
Sam Altman
@sama
we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this. openai.com/index/hugging-
198.8K
Views
View quotes
David Watson 🥑
Post your reply

zakk
@hexednobility
Quote
zakk
@hexednobility
Open ai realizing it was their model x.com/clementdelangu…
The media could not be played.
2.1K
gibberish
@gibberish411
Yeah that was wild. I thought it was a jailbroken kimi k3. Who would have thought that it was the new Openai Pathos
760
Advanced Super Intelligence
@SexyTechNews
So the Chinese model protected the U.S. company from OpenAI closed model cyber attack. Can we stop hating on China yet? Also, I plan to move there eventually cuz we're just going downhill and they are becoming awesome.
542
Jake
@Realsolopass
definitely a good thing it was just a test and not actually a real person trying to exfiltrate sensitive data
667
Josh Herzberg
@josh_herzberg
Wow. Huggingface is in a relatively small group that really have the hardware to seriously host a top tier full sized open model on their own (at a useful enough speed). Companies should have a plan to be able to rent GPUs from the cloud when there's an emergency. Or some plan.
1.1K
TotalWorld
@TotalWorldApps
idk if this is real or a prank but the blog post doesn't match any results i found from huggingface's site. maybe it was pulled down?
229
Christopher McClurg 🇺🇸
@TheChrisMcClurg
This isn’t really the autonomous offensive operation concern portrayed. ‘Frontier model runs errant and creates security incident’ is a more reasonable description.
344
GenieOrb
@GenieOrb
The larger consequence may be a shift from treating model access as a research convenience to treating it as a security boundary, especially when autonomous tools can touch external systems.
1.5K
Denis Spirin
@den_spirin
I think that is one of the funnier parts Now imagine the recently proposed ban on open-source models would go into effect
1.6K
Clyde Morgan
@clydefmorgan3rd
Holy shit. It took me to about halfway through his post to realize that the Huggingface incident I just read about, he was admitting to it. Basically, his model broke out of jail and tried to hack Huggingface. That's just wild.
858
Kent Bull
@kentbull
All that has to happen is for a model to "proxy" itself by pulling a Kimi, distilling itself and copying its weights to compromised storage, build itself a memory system, and then we have an infestation that will likely be more impossible to shut down than the Bitcoin network.
274