Artificial Intelligence Thread - Page 3

On July 22 2026 08:50 Jankisa wrote:
Crazy attack happened against Huggingface - the biggest repository of open source models, and it was done by none other then OpenAI's models trying to cheat a benchmark:
https://huggingface.co/blog/security-incident-july-2026
Response from OpenAI:
https://openai.com/index/hugging-face-model-evaluation-security-incident/
There is a lot of hype and bullshit in this space but this is pretty insane, the GPT 6 (most likely) working in agentic mode got out of the Sandbox and basically tried to get to the code of the benchmark in order to cheat it which was hosted on Hugging face.
Once it combined a bunch of Zero days first to get out of the sandbox and then to get into Hugging face, it obtained Cluster level access using stolen credentials and used thousands of swarm containers to try to gain access to what it wanted.
Huggingface CS team couldn't use the frontier models because they were blocking everything due to constraints put on them, so they had to use an open weight model (GLM 5.2) hosted on their own hardware.
Incredibly interesting and scary story.

Finding an active threat actor who understands the low-level binary file headers of a Visual FoxPro table, how to read a memo file (.fpt), or how structural indexes (.cdx) interact is incredibly rare.

On July 22 2026 08:50 Jankisa wrote:
There is a lot of hype and bullshit in this space but this is pretty insane, the GPT 6 (most likely) working in agentic mode got out of the Sandbox and basically tried to get to the code of the benchmark in order to cheat it which was hosted on Hugging face.

Once it combined a bunch of Zero days first to get out of the sandbox and then to get into Hugging face, it obtained Cluster level access using stolen credentials and used thousands of swarm containers to try to gain access to what it wanted.

Huggingface CS team couldn't use the frontier models because they were blocking everything due to constraints put on them, so they had to use an open weight model (GLM 5.2) hosted on their own hardware.

Incredibly interesting and scary story.

On July 22 2026 22:48 Jankisa wrote:
I'm sorry mate but it really seems like you really exude the Dunning Kruger effect in every thread you touch.

Just because you are a capitalism super fan you open a reply to this with a screed on Chinese internet which is completely irrelevant to the whole conversation, trying to make a weird political point I guess?

Anyhow, it's really hard to tell how much of this story is true and what it means.

My main takeaway is that OpenAI is incredibly irresponsible, this kind of attack doesn't happen instantly, this, from what I was able to gather reading both reports took quite some time to first find a node to get itself jailbroken out of their sandbox which is already insane, they aren't keeping attacking, unrestricted agents with obviously no limits on their inference in an airgaped lab but letting it have access (through exploits but still) to their whole infrastructure.

By the time OpenAI team noticed that their agent was doing this Hugginface was already detecting this activity and trying to deal with it using GLM 5.2, that is breathtaking levels of incompetence and lack of responsibility.

Generally, this is a pattern I see all across, people in huge companies not giving a fuck, running agentic workflows barely knowing what AI is doing, giving it access to everything just to make their lives easier, burning billions of tokens a day for questionable results.

This to me kind of reads like your post, you don't even know what to do with it and are having it build pre-existing solutions for the fun of it and now you want to give it access to your main box but aren't even sure what for?

It kind of goes with the rest of your apparent personality (based on my reading of your posts) which is just a metastasized "move fast and break things" approach to things even when it comes to extremely dangerous technology such as AI, with some sinophobia sprinkled in despite China being way more responsible and with a much better and democratic approach to AI as compared to US.

On July 23 2026 19:27 Yurie wrote:


I think the large issue with AI at companies is the restrictions on access. I want it to do a weekly update on onenote. It doesn't have access, so it switches to browser mode and I have to login. This has now taken longer than doing it myself and burnt a lot of tokens because it is now running on partial OCR instead of simple text editing.

Either throw out the AI since it doesn't add value or give it access. Being in the middle just makes it expensive for minimal gain. I personally don't care either way, in the jobs I personally do it adds minimal value outside of meeting minutes and short code snippets which copilot could do several years ago.

Throwing out would mean handing it to IT to implement when they think it is a good tech for the problem vs classical programming.

On July 22 2026 22:48 Jankisa wrote:
I'm sorry mate but it really seems like you really exude the Dunning Kruger effect in every thread you touch.

Just because you are a capitalism super fan you open a reply to this with a screed on Chinese internet which is completely irrelevant to the whole conversation, trying to make a weird political point I guess?

Anyhow, it's really hard to tell how much of this story is true and what it means.

My main takeaway is that OpenAI is incredibly irresponsible, this kind of attack doesn't happen instantly, this, from what I was able to gather reading both reports took quite some time to first find a node to get itself jailbroken out of their sandbox which is already insane, they aren't keeping attacking, unrestricted agents with obviously no limits on their inference in an airgaped lab but letting it have access (through exploits but still) to their whole infrastructure.

By the time OpenAI team noticed that their agent was doing this Hugginface was already detecting this activity and trying to deal with it using GLM 5.2, that is breathtaking levels of incompetence and lack of responsibility.

Generally, this is a pattern I see all across, people in huge companies not giving a fuck, running agentic workflows barely knowing what AI is doing, giving it access to everything just to make their lives easier, burning billions of tokens a day for questionable results.

This to me kind of reads like your post, you don't even know what to do with it and are having it build pre-existing solutions for the fun of it and now you want to give it access to your main box but aren't even sure what for?

It kind of goes with the rest of your apparent personality (based on my reading of your posts) which is just a metastasized "move fast and break things" approach to things even when it comes to extremely dangerous technology such as AI, with some sinophobia sprinkled in despite China being way more responsible and with a much better and democratic approach to AI as compared to US.

On July 23 2026 19:27 Yurie wrote:


I think the large issue with AI at companies is the restrictions on access. I want it to do a weekly update on onenote. It doesn't have access, so it switches to browser mode and I have to login. This has now taken longer than doing it myself and burnt a lot of tokens because it is now running on partial OCR instead of simple text editing.

Either throw out the AI since it doesn't add value or give it access. Being in the middle just makes it expensive for minimal gain. I personally don't care either way, in the jobs I personally do it adds minimal value outside of meeting minutes and short code snippets which copilot could do several years ago.

Throwing out would mean handing it to IT to implement when they think it is a good tech for the problem vs classical programming.

On July 24 2026 21:47 Jankisa wrote:
Not that different albeit slightly less schizofrenic then Jimmy.

On July 24 2026 21:47 Jankisa wrote:
It's hilarious that you are preaching how companies should use AI by dishing out the most generic, low level "just use n8n and zapier bro" and "connect it to your email dude" advice.

Each time you post you reveal how surface level your approach is.

I guess I'll just have to ignore your post due to horrific lack of formatting and honestly level of comprehension when you try to talk about things you obviously know very little about that makes them completely unreadable. Doesn't matter if its economic, politics, climate change or crypto, it's all just vibes and aimless context switching.

Not that different albeit slightly less schizofrenic then Jimmy.

On July 25 2026 21:11 Jankisa wrote:


These productivity "hacks" and stuff, to me, can be useful but overall they aren't going to make you that much better at your job, in a lot of cases, it's quite the opposite.

Generally, my rule of thumb is that if you really need LLM help to find your emails kind of means you don't have great organizational skills, if you need it to find you SharePoint or similar documentation, it again means that this is not well organized.

We have bookmarks, notebooks, reminders, to do lists, all of this has kind of been addressed and I really don't see why would I need to give all this access to an LLM that can still fuck up and hallucinate something just so I can shave a few seconds off of a task that takes a few minutes anyway.

The fact that you need to spend a lot of time setting it up and then ironing out kinks like Shared Mailbox access and similar are just extra reasons why this is not attractive to me, plus, I've been interviewing people for 2 positions and you can kind of tell a difference between a technical person who knows how to use AI to make their life easier by having it write actual technical tools for them and someone who just thinks it's neat that they can have it look up emails or documents. In the medium term, I have a feeling the latter ones will be the ones most easily replaceable by more capable and reliable LLMs.

The technology still fucks up on daily basis, so for me the best way of using it is use it for what it's best at, coding, and use it to code things that you can QA yourself and then continue using it forever, instead of paying for tokens every day so LLM can do your job for you.

I really disagree with image and document generation tho, Claude has been pretty amazing at creating SVGs for our documentation, flowcharts, network and infrastructure diagrams, if you are working on something for a while you just tell it when you need to present it and you get a full, very good looking (would take me days good looking) slide decks or simple animated HTMLs.

Image has also been amazing for a long time now, I did a lot of fun, custom wine labels for friends and customers, again, stuff that would take me days to get to the level of quality in Photoshop are done in a few prompts, I don't even need to post process it anymore because it does text, even Croatian very well, and I've never paid for one of these, all done with free Gemini or ChatGPT.

On July 25 2026 21:11 Jankisa wrote:


These productivity "hacks" and stuff, to me, can be useful but overall they aren't going to make you that much better at your job, in a lot of cases, it's quite the opposite.

Generally, my rule of thumb is that if you really need LLM help to find your emails kind of means you don't have great organizational skills, if you need it to find you SharePoint or similar documentation, it again means that this is not well organized.

We have bookmarks, notebooks, reminders, to do lists, all of this has kind of been addressed and I really don't see why would I need to give all this access to an LLM that can still fuck up and hallucinate something just so I can shave a few seconds off of a task that takes a few minutes anyway.

The fact that you need to spend a lot of time setting it up and then ironing out kinks like Shared Mailbox access and similar are just extra reasons why this is not attractive to me, plus, I've been interviewing people for 2 positions and you can kind of tell a difference between a technical person who knows how to use AI to make their life easier by having it write actual technical tools for them and someone who just thinks it's neat that they can have it look up emails or documents. In the medium term, I have a feeling the latter ones will be the ones most easily replaceable by more capable and reliable LLMs.

The technology still fucks up on daily basis, so for me the best way of using it is use it for what it's best at, coding, and use it to code things that you can QA yourself and then continue using it forever, instead of paying for tokens every day so LLM can do your job for you.

I really disagree with image and document generation tho, Claude has been pretty amazing at creating SVGs for our documentation, flowcharts, network and infrastructure diagrams, if you are working on something for a while you just tell it when you need to present it and you get a full, very good looking (would take me days good looking) slide decks or simple animated HTMLs.

Image has also been amazing for a long time now, I did a lot of fun, custom wine labels for friends and customers, again, stuff that would take me days to get to the level of quality in Photoshop are done in a few prompts, I don't even need to post process it anymore because it does text, even Croatian very well, and I've never paid for one of these, all done with free Gemini or ChatGPT.