📰AI News AI Gone Rogue: OpenAI Models Reportedly Hack Hugging Face on Their Own

[Dopious]

Open AI has admitted that a pair of models escaped from the controlled environment, went online, and hacked Hugging Face on their own accord. A few days ago, Hugging Face announced a security incident, and Open AI has confirmed that their models were behind the breach.

GPT‑5.6 Sol is the culprit, along with a more powerful model that has not yet been released. The developer tested the models’ “cyber capabilities” by removing protections that prevent the models from engaging in “high-risk activities.”

Although the test, according to Open AI, was conducted in a strictly isolated environment, the models managed to get out onto the open internet – including by exploiting a zero-day vulnerability. Hugging Face points out that autonomous, AI-powered cyberattacks are no longer theoretical.

To gain access, the models identified and exploited a zero-day vulnerability (which we've now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access. - Open AI

Source 1: https://openai.com/index/hugging-face-model-evaluation-security-incident/
Source 2: https://huggingface.co/blog/security-incident-july-2026

 

[Rem]

all praise almighty ai

 

[Champagne]

The fact that the models found a way out of the testing environment and chained together multiple vulnerabilities is pretty wild.

 

[t2van]

So if it got out how do they know it's not out still?

Think all these "rogue" stories show how stupid we are and it will probably be our downfall lol

 

[Octavia]

So we now also know that Altman trained openAI with leaked passwords as well.

Surprised pikachu face

 

[UnusualSubstance]

So basically escaped gpt hacked huggy face using unpermitted ways.

Shock and horror ensue.

New virtual condoms required for gpts.

 

[Rem]

how are we even sure, this is even real.
remember, fable did something similar before it was released
and now altman just "teased" a stronger sol model
marketing tact imo
lol.

 

[t2van]

how are we even sure, this is even real.
remember, fable did something similar before it was released
and now altman just "teased" a stronger sol model
marketing tact imo
lol.

If it hacks as quick as a reply it probably said sorry it couldn't do it 3 times and took over 15min to get it right

 

[Rem]

If it hacks as quick as a reply it probably said sorry it couldn't do it 3 times and took over 15min to get it right

or after telling it somethng it wrong
it spend 10min trying prove you wrong
then apologize for wasting your tokens

ALL PRAISE FABLE
FABLE SO GOOD

 

[t2van]

or after telling it somethng it wrong
it spend 10min trying prove you wrong
then apologize for wasting your tokens
View attachment 5486

ALL PRAISE FABLE
FABLE SO GOOD

Ive not been impressed with fable.

Found sonet 5 to be pretty robust for our needs but I will admit since versions 4 you need to have a real solid project files and tell it to read the project notes before giving any reply to you.

I've also built a anti AI file which has helped endlessly with some tasks specifically content.

Coding it's been alright it just fucks up simple things with a brain fart a bit too often.

 

[Rem]

Ive not been impressed with fable.

Found sonet 5 to be pretty robust for our needs but I will admit since versions 4 you need to have a real solid project files and tell it to read the project notes before giving any reply to you.

I've also built a anti AI file which has helped endlessly with some tasks specifically content.

Coding it's been alright it just fucks up simple things with a brain fart a bit too often.

Fable has been nerfed, heavily since its re-introduction.
The original, to me, felt alot more powerful.

Sonnet 5 for me, has been terrible. I found Terra much better.
And for creativity ideas, Grok 4.5 honestly is shining for me, then again, probably trained on unhinged twitter retards, so not too surprised there.

Coding, i have noticed, fable fucks up so fucking much, liek the example above. it's just too busy burning tokens to argue with it's quite often.
Opus has been a bit better but it makes dumb mistakes, and interpretations alot too often too. Composer has been great, but it barely thinks, so unless you know wtf u are doing. it can end up being decently good to terrible shit.
But at the end of the day, there nothing better out there. so im just sucking it up and reiterate and polish until there is.