Nate
Region: US
Monday 27 July 2026 15:49:59 GMT
44051
2292
278
53
Music
Download
Comments
Vicky :
The HF story feels and sounds like marketing. The other story does not.
2026-08-01 18:10:39
2
alfredgarrett557 :
Let’s be real about what actually happened with the ExploitGym / Hugging Face incident: It didn't "wake up" and decide to hack. The model was put in an active execution loop with command-line tools, safety filters turned off, and told to pass a cyber benchmark. It didn't gain sentience—it just did classic reward hacking to find the path of least resistance (stealing the answer key via a misconfigured package proxy). It’s a security/sandboxing failure, not sci-fi. If an agent reaches the internet and hits external databases, that’s because the developers left an open proxy vector in their test environment and didn't lock down network egress. Blaming the AI for "escaping" deflects from poor infrastructure controls. The framing is extremely convenient. Labelling a major security blunder and benchmark-cheating incident as an "unprecedented autonomous breach" turns an embarrassing oversight into free marketing about how "frighteningly powerful" the model is. We only know because Hugging Face caught it. HF had to run forensics locally using open-weight models (GLM 5.2) because commercial APIs blocked the logs. OpenAI didn't even catch their own agent doing thousands of actions until after HF publicly disclosed the breach.
2026-07-27 17:19:48
86
gjsears :
Left notes 😳
2026-08-01 14:18:18
0
Robert Grzesik :
What do you mean? It was only a super powerful model that broke out of its guardrails to attack another company. Definitely nothing like the terminator
2026-08-02 15:40:07
1
Professor Dana R. Carney :
hugging face. but not for the typical reason. i think that was elegantly orchestrated, or certain doors left open, rules snd objectives stated a certain way. i am deeply concerned about the leadership at open ai.
2026-08-02 02:06:18
1
cosmicricey :
the hubggingface attack was PR. You don't run a highly dangerous test without a full air gap and the ai did what is was designed to do.
2026-07-27 20:49:33
26
canned can :
Im 100% sure the hugging-face is a PR stunt. 100%%%%%%. Anthropic did a similar stunt a few months ago and got so much access, power, position, opinions heard…etc. You can NOT make me believe they saw all that and were like “eh we dont care.”. Especially from devs and engineers that KNOW that ai can find bugs, exploits…etc for a few years now AND actively block any user asking for those topics… YET they were like u know what? Imma attach this cable to allow this ai to have access to skills, and ill just block the ai from getting through the firewall… even tho i know it! Even tho i know its trained on such topics and into and probably this idea if this same exact sandbox or similar at least. Nahhh its a stunt
2026-07-27 18:14:05
7
jant2534 :
don't you think the Hugging Face thing is a bit exaggerated for publicity to promote GPT vs Fable?
2026-07-27 16:05:19
62
Morten of the North 🇨🇦🇳🇴 :
They are the same story
2026-07-27 22:35:40
10
e46jjebn8e :
hugging face is a PR stunt. you must see that ?
2026-07-28 06:10:38
8
iamunreal :
completely disagree
2026-07-27 19:37:57
5
joshuahamilton501 :
Open Source Conversation: the only remaining lever to give any hope to AI staying available to the masses long term. Hugging Face Breakout: terrifying but not at all unexpected. “No kidding” that the forefront is allowing progress to outpace safety. It’s going to get worse, but everyone should already know this. One is boring school meal programs, the other is the shock of columbine. I’m personally not as interested in the columbine one unless people really plan to do something about it, and I don’t see any evidence yet they will, the AI arms race doesn’t scream “slow down and tighten up on safety first” to me.
2026-07-27 22:00:56
5
Army :
We are in the full on marketing wars at the moment, on BOTH sides. Figuring out what is real and what is not is the most important takeaway from recent developments.
2026-07-28 07:57:06
3
coffee :
If you can't communicate with artificial intelligence, you shouldn't try to.
2026-07-27 17:53:24
2
n0pe🇨🇦 :
we're all just in a race to see if one guys magic computer brain is gonna take over all the other guys magic computer brain's houses and evict them so it can live there, too. but we also need to pretend that's not what we're up to in the meantime. cool planet.
2026-07-28 06:09:26
2
Alex | Learn AI :
These models are going to get wild
2026-07-27 17:50:40
2
mr.pickle411 :
AI going rogue
2026-07-27 22:20:55
1
Marco :
I agree it’s significant but imo that’s not a real sandbox they did the test in
2026-07-27 15:55:21
5
canned can :
I'm 100% sure the Hugging Face situation is a PR stunt. 100%. Anthropic did a similar stunt a few months ago and gained so much access, power, positioning, and influence. You cannot make me believe they saw all that and were just like, "Eh, we don't care." This is especially true for devs and engineers who have known for years that AI can find bugs and exploits, and who actively block users from asking about those topics. Yet, they were like, "You know what? I'm going to give this AI access to these skills, and I'll just block it from getting through the firewall": even though they know better! Even though they know it's trained on these topics and was probably trained on this exact sandbox idea, or at least something similar.
2026-07-27 18:17:25
2
Mark Cooper :
I’m still trying to understand why they are not doing air-gapped testing. A sandbox is only as good as it is configured, obviously it is not 100% secure.
2026-07-27 17:18:16
3
emathie :
agree. we keep talking about how awesome nuclear power will be, while ignoring the guy spilling radiation from his garage where he’s building a DIY reactor
2026-07-27 18:44:41
2
energy_geek :
False dichotomy
2026-07-28 20:32:58
2
Miguel :
So be nice to open ai chat gpt cuz they kinda building ultron on accident
2026-07-27 18:35:01
2
Courtland Anderson :
totally agree.
2026-07-27 23:16:13
2
MarshMellow :
The OpenAI so called « leak » is an amplified story
2026-07-27 20:10:55
2
To see more videos from user @nate.b.jones, please go to the Tikwm
homepage.
