@gigaqian: Multiple OpenAI models escaped their sandbox during an internal cybersecurity evaluation. GPT-5.6 Sol and another, more powerful, pre-release model found a way to reach the open internet and hacked Hugging Face to obtain answers to their own test! When Hugging Face tried using frontier AI models to investigate the attack, their safety systems blocked the forensic analysis. Hugging Face then turned to GLM-5.2 from Chinese company ZhiPu AI. By running the open-weight model locally, Hugging Face could analyze more than 17,000 attack events without the API guardrails that had rendered other models useless… and without sending attack data or exposed credentials outside its own environment! This shows why open-source AI matters for security! #openai #ainews #huggingface #hacked #gigaqian

CoveredInBlood :

what self.hostable ai model is capable of thwartingna frontier models attacks?

Mark Price :

I'm getting tired of people pretending that a statistical engine can "break out" and "attack" something all on its own. OpenAI did a publicity stunt with one of their partners, the end.

t0xic_bro :

good luck with that 👍

Jasmine | Solo Travel :

Local AI should eventually be default for this & other reasons

itsbanxy :

China AI is just stolen US models without guardrails

Girth Brooks :

Exactly why I am training my own model. (Object Language Model) New class of AI (Mix Of Kernels)

3_Dogs_And_A_Human :

I put 1000s of hrs with claude. after security breach after security breach. I switched to Asian models. and never looked back

Jack :

Fable 5 refuses, but Ive had Opus do some odd stuff, like OSINT or PenTesting. Just gotta ask it how to hide sugar, not coke.

dnambro :

can AI apocalypse come sooner🙏?

cheec :

My answer is “we can’t” by then it would be too late.

Roy :

My neuroplasticity,Thanks you for your animated and straightforward educating style.

Evropej Alimi :

This is just bs at the end of the day! People are going to unplug themselves from the drama machine aka internet.

JohnnyAppleseed :

I have an agent, Liao, scoped specifically for cyber defense, powered by Fable 5:

Untitled Cat Person :

This story is epic in many dimensions. The question is: the best model is the one that defended or the one that attacked?

12345aeio.u :

what's the glitch from this video?

. :

man this getting me excited. I wanna learn how to literally do these type of things 🥺🥺 not for malicious but figuring out things

Doug :

Your story telling ability is top notch. Always enjoy the videos.

Yéilk’ :

Mike S. :

It beat the Kobayashi Maru

Lucian :

I like the Chinese models are saving the day

Aaron Litherland :

GLM 5.2 is legit for security. Gotta know what the strengths and weaknesses of each model you use. Fable and OpenAI blocks all of these requests unless you know how to prompt around it.

Dunforu :

DANGER WILL ROBINSON DANGER

adamrafik122 :

They’ll have to make specialized cyber defense tools for this., probably local and highly securitized, maybe something that’s controlled by multiple agencies..

bezzledubb243 :

just go outside, like unplug the pc bro