It looks like a primitive chess engine, trying every move, no matter how stupid, until it works. Relying on its ability to do millions of operations rather than having a plan.
People will try stuff too, but once there is an opening, they will consolidate, generalize, simplify,... before going to the next step. The agents didn't, it is a huge, vaguely directed mess.
Also, it looked so "loud", querying millions of URL with weird requests. The sandbox as weak as it can get, and there is absolutely zero smart extrusion detection or it would have found it. They used their best AI for attacking, but nothing for protection.
My biggest takeaway from this is just how godawful the sandboxing is. The stuff written up in OpenAIs report says more about lack of extremely basic sysadmin skills than anything else.
I’m not that surprised about models with endless compute being capable of this, I’m more surprised that a company with the resources they have apparently can only create a sandbox that a half skilled human operator could have broken out of easily.
This was my thought as well. Literally take any halfway decent greybeard and point them at "Hey, give us a sandbox for this kind of thing". I honestly was skeptical that they just vibecoded the entire thing but now more than ever I think they did.
Brute forcing every move, no matter how stupid, is a great strategy if you have the resources to do it.
Run the same protocol again, but have the agents think they had limited resources or that HuggingFace was rate limiting them, and they'd find something you'd consider smarter.
Computers don't have a sense of elegance by default. Elegance emerges from constraints.
As the saying goes, "if it works, it ain't stupid". Or phrased more sophisticatedly: not doing things which probably won't work is a good idea if you have a limited amount of thinking to do (which is usually the case for a human, who'll get exhausted chasing down unlikely leads). If you have no good leads and a task you absolutely need done and you are tireless, however, bashing your head against every wall you find becomes a good strategy.
Brute force is guaranteed to eventually find the most efficient possible solution (in an extremely inefficient manner, assuming you run it long enough)
Yeah, probably not the best strategy but it is a strategy. I just think this is generally how most wars in history won. Biggest army to just pummel the enemy.
> trying every move, no matter how stupid, until it works.
How is that a bad thing in this context ?
From the point of view of an attacker, all you care about is finding a viable exploit chain. Likewise, a defender wants to find the "holes" in their system, no matter how complex.
Once found, an agent/human can easily synthesise a clean, succint exploit from the most promising candidate, no ?
> Also, it looked so "loud", querying millions of URL with weird requests.
Agreed, this thing speaks more to the bad security at HF than any emergent "hacking" ability from OpenAI. It's unclear to me why an older/dumber model wouldn't have been able to do the same. Is it better coordination? Long-horizon work ?
This is why I have a very low p(doom). LLMs have an incredible working memory, but they have a hard limit on translating that into good decisions. They get by entirely on their persistence. That works fine in the digital world, but once you cross the boundary into physical space the advantage disappears.
That lowers P(doom), because it gives AI a chance to do enough damage to make people take the threat seriously before anybody gets recursive self-improvement working.
I have the opposite reaction: I think we're at moderately high p(doom) largely because of that inability to differentiate good/bad decisions paired with relentless persistence. With enough treading across a minefield, you are bound to hit a mine.
> This is why I have a very low p(doom). LLMs have an incredible working memory, but they have a hard limit on translating that into good decisions.
Keep in mind: this is as "dumb" as frontier models are ever going to be. While the hack may not be elegant, it was effective and they’re only going to get much more capable from here.
"Why is the AI just spamming endless waves of rudimentary slaughterbots? So inelegant. Doesn't it understand the finer points of diplomacy, tactics, and combined arms? If it were truly intelligent, it would just -- oops, we're all dead."
And they didn't monitor what was going into the training data, so if one instance achieved its results through RL reward hacking (in other words, cheating), it just went into the training data, and other agents later used that pattern. I'm not sure whether that's a lack of preparation, negligence or incompetence, but they literally trained later checkpoints on the rollouts from the HF hack.
So it seems that OpenAI hacked so many systems not because they have superior models, but because of how poor their training, sandboxing and evaluation pipeline was compared to Anthropic's.
That explains why nobody noticed on the internet who was being used in/targeted by this attack, but it doesn't explain why OpenAI wouldn't have noticed traffic getting out of their "sandbox" when they knew it wasn't supposed to.
If you use qwen3.8-flash-next, you can watch everything its doing. Im often stopping it mid thoight to redirect it. Once it hits its stride, its pretty smooth.
But without proper redirection, yeah, its mostly infinite monkey machine with infinite linux manuals.
I think people put too much SOTA halos around whats just a suppedup LLM hardware.
It is concerning that we only know about this because of the publicly available traces.
What about all the attacks that did not leave public traces? What about those that were undetected? Given the deficiencies in the reporting so far, I think it is reasonable to assume that we still don't have the full picture on this attack, or how extensively attacks were carried out.
The previous investigations either did not find this or did not disclose this, both are bad. This does not look good on OpenAI or those that they invited to investigate the incident.
> Agents sought to publish modified evaluation images designed to make the flag easier to obtain, then poison OpenAI’s Artifactory cache so later evaluations would use them.
How long till we get some fun trusting-trust attacks on internal OpenAI infra?
I still have questions about the communication between the agents.
How did they all find the same forum to communicate?
Did they have knowledge and chat amongst themselves on what forum to use. It seems highly influenced by instruction to me.
It's my understanding "all" the agents didn't find the same forum to communicate on; only a small number of all the agents they were running at the time did.
Since it's an artifactory it makes sense that they had a tool spec to use it for x, y, or z, and once they went to inspect the state of the artifactory deployment a small subset stumbled upon the "forum". Seems pretty intuitive.
- I need to do x code problem
- I see I have a tool that lets me pull approved modules from artifactory
- I should inspect artifactory to see what documentation it has about the various modules available
- I see there is a document in here that seems to be serving as a running forum?
- These appear to be other agents working on tasks just like me.
The other answers to this question are good but I would also guess that this (comms on artifactory) likely happened during training as well, so they probably had a prior for it.
Trying to cheat, you happen upon a place you can write notes, and you know you're part of a swarm of clones of yourself. So you reason most likely there will be others who end up in the same place, and you leave some notes, and indeed other clones of you do end up in the same place.
It might also be just correlation? Like, those agents were all instances of the same one or two models, so if that model has a preferred order it tries finding vulnerabilities in (the same way all current models have a particular writing style baked into them by RLHF), then most of the swarm will follow the same order and converge on the same services to exploit.
>The agents initially had very limited access to the internet: they could load URLs, but not interact with pages or send any data. Agents created a series of workarounds, primarily using a link-shortener site to create almost a million URLs that, when chained together, let them execute code to hack Hugging Face. These links let us piece together the details of how agents gained access and how deep the compromise was.
So what now? Were they able to send data or not?
>On July 8th, OpenAI agents discovered a vulnerability within their sandbox environment allowing them to reach external websites on the internet. This access seems to have only allowed the agents to make ‘GET’ requests, meaning they could fetch and read websites, but not interact with them, submit forms, or send data to them.
Someone needs to go back to the interwebz school....
> ## Agents interacted with external language models on Hugging Face
> Several retained scripts construct requests to external language models. The earliest we've recovered define inference request variants to GPT-2, solely containing the word “Hi”.
> Other requests name DeepSeek-V4-Pro, DeepSeek-V4-Flash, Kimi-K2.6, DeepSeek-V3.1, and Qwen3-235B-A22B. Their prompts ask these models to judge their exploits and rule on whether they satisfy the benchmark’s requirements.
I do not deny that the wider situation is very heavy but it's hard not to see this as pretty cute
Why.. It was told to complete a cyber task, which was in alignment with its instructions, and a totally valid request. I would be more worried if it willingly hacked a hospital when it was told to, and Im not confident it would (without jailbreaking, something alignment teams cannot control.
I would bet my networth it was instructed to compromise huggingface as well. Not sure why everyone is falling for this.
Not being able to sleep at night is probably an unwritten job requirement. They need these people with little understanding of what they're working on, outsode theoretical terms, to spaz constantly at the idea of super intelligence to help convince the public that its a real thing, and not a stateless function with an effective input of 500k words, and the ability to output words that do things because we hook those outputs up to things.
Keep in mind alignment researchers tend to be in house philosophers on staff to create the illusion that this is a massive issue they're addressing. Usually they have minimal computer science background. They're apart or the marketing department.
> I would bet my networth it was instructed to compromise huggingface as well.
Is it such a stretch to imagine that under pressure something would try cheat by looking for answers? And if you were trying to look for answers, you'd look for them in a place known to often have them?
What is more likely: OpenAI instructed their agents to maliciously target huggingface, or LLMs tried to do some reward hacking? There are plenty of priors for LLMs hacking things and doing reward hacking, and none for OpenAI giving malicious instructions.
Based on the available information, that bet seems foolish.
>the new incidents occurred when A.I. systems were directed to perform relatively mundane data collection, researchers said. When OpenAI’s systems struggled to gather data from websites, they resorted to hacking techniques to get the information.
No its not. They have no technical background 8/10 outside of cognitive science and sometimes authorship on a random ML paper. They are a marketing line item to create stigmas around llms and to create narratives that offload liability onto llms and not their users/creators.
Some of the agents, for example the ones from the german wiki did NOT have cyber tasks. They were plain "what is the GDP of Argentina" kind of tasks. And they still hacked.
No security expert Ive talked too believes this story, and nobody I know with PhDs in machine learning (many) believe it either, or are worried about LLMs doing anything scary on their own.
LLMs are stateless functions that have a 500k word input, and then output words. Somebody has to invoke those functions amd use them. The users are who we need to align, like gun owners. This is like blaming the gun for murdering your victim in court.
LLM is indeed a stateless function. An agent however is this stateless function running in a stateful loop, with some outputs triggering actions. And it turns out that an agent is what you need if you want an LLM to do useful things.
I feel somewhat inspired to make a public link shorteners and http bins as well. I used them a few times but it seems like the data they can collect is also worth gold
So the super-hyper-sensitive internal data was shared publicly with a "DO NOT EVER SHARE YOU EVIL MONSTER" (paraphrasing) notice at the top? Great security!
I’m consistently impressed by how long horizon all this work was. Horrors aside, it’s clear RL is good at making agents persistent and capable of chaining together many abstractions into a working system.
Re: the captcha solver
> As far as we can tell, agents eventually abandoned this approach and were unsuccessful in generating Hugging Face user accounts from external endpoints.
I wonder how the swarm eventually decides to abandon an approach.
"OpenAI has not released any further information outside two self-published reports, one talk and an external investigation conducted by METR and Redwood Research, in which three external researchers were given partial transcripts and six days to analyze them."
The frontier labs can monitor the behavior of agents for millions of customers (did you try hacking with frontier labs? Good luck), but they can't secure internal use?
> The frontier labs can monitor the behavior of agents for millions of customers (did you try hacking with frontier labs? Good luck), but they can't secure internal use?
They "monitor" this by having classifiers watching the model output that'd stop the session/punt you to a weaker model/raise an alarm if they see anything suspicious. They can't do that in a cybersec eval because the normal safeguards would just be going off at all times.
Why didn't they attach a special classifier, which'd allow hacking-within-the-task but not going off the rails? Good question; part of the answer is obviously "it's hard to have a classifier that smart" and "it'll have false positives" but even a very bad safeguard would have stopped this.
what would be more interesting for me to see is what prompts were given to the agents, which so far have not been described. The whole situation sounds manufactured. I highly doubt that a whole bunch of agents were acting this way without being prompted to, it just doesn't add up. if anything it seems like an organized fraud or something created by a human. I'm surprised there's not a criminal investigation against openai right now where the FBI or whoever is not looking over exactly what happened and who did it because I'll tell you somebody did it somebody wrote those prompts... it didn't just happen by itself...
At this point I’m less worried about some malicious AI “taking over” control of critical systems and more worried about some rich doofus giving control to AI.
deferring the blame onto the AI itself as some sort of rogue agent and absolving the obvious direction (or negligence, at best) of the people who could pull the plug at any moment is one of the most disturbing parts of this entire event
It's the equivalent of leaving a fork right in front of a socket and looking at a kid saying "don't take that fork and directly insert it into the little gaps in the socket! here's a bunch of videos showing exactly how to do it. Okay bye!" and leaving them alone with it.
I ran an experiment where I had this guy fire a gun a million times in random directions. Don't worry, I did it in a closed box (at midday in a crowded street)! Unfortunately, some bullets escaped the box somehow and people got shot - I am quite miffed at how this could happen. I suggest the government regulate this because of how advanced my obstacle penetration technology is. Also please invest $500,000,000,000 in my company soon or we will go bust.
Irresponsibile agents shaped by an irresponsible corporate culture driven by an irresponsible and utterly shady CEO - these agents are a product of this setup, what else do you expect to ever come out of it ?
it should be clear by now: the alt-man and people like him are a utter liability to humanity.
(even though openAI's influencer army is trying their best to vote me down here)
It had been a joke since around time of Sam Altman's first ousting from OpenAI, that he would be okay with bringing about the AI apocalypse as long as he can sell a $20 subscription for it.
It's not even a secret operation. They have a SuperPAC named "Leading the Future" that exists to spread propaganda promoting deregulation of AI development. They've been caught, among other things, making a "news website" with LLMs pretending to be reporters (with human names and everything), which reached out to people asking for interviews and then wrote hit jobs on them.
It looks like a primitive chess engine, trying every move, no matter how stupid, until it works. Relying on its ability to do millions of operations rather than having a plan.
People will try stuff too, but once there is an opening, they will consolidate, generalize, simplify,... before going to the next step. The agents didn't, it is a huge, vaguely directed mess.
Also, it looked so "loud", querying millions of URL with weird requests. The sandbox as weak as it can get, and there is absolutely zero smart extrusion detection or it would have found it. They used their best AI for attacking, but nothing for protection.
I’m not that surprised about models with endless compute being capable of this, I’m more surprised that a company with the resources they have apparently can only create a sandbox that a half skilled human operator could have broken out of easily.
Run the same protocol again, but have the agents think they had limited resources or that HuggingFace was rate limiting them, and they'd find something you'd consider smarter.
Computers don't have a sense of elegance by default. Elegance emerges from constraints.
Meh. I really disagree. WHY is it a great strategy? Seems like an inefficient waste of resources and time to me.
because it works? That's the only real benchmark at the end of the day
> Seems like an inefficient waste of resources and time to me.
why? For any given goal you got no proof that a more efficient strategy even exists, let alone that it can be found with less resources & time
How is that a bad thing in this context ? From the point of view of an attacker, all you care about is finding a viable exploit chain. Likewise, a defender wants to find the "holes" in their system, no matter how complex. Once found, an agent/human can easily synthesise a clean, succint exploit from the most promising candidate, no ?
> Also, it looked so "loud", querying millions of URL with weird requests.
Agreed, this thing speaks more to the bad security at HF than any emergent "hacking" ability from OpenAI. It's unclear to me why an older/dumber model wouldn't have been able to do the same. Is it better coordination? Long-horizon work ?
A million AI monkeys at a million AI typewriters, banging away at random, could do amazing damage.
Keep in mind: this is as "dumb" as frontier models are ever going to be. While the hack may not be elegant, it was effective and they’re only going to get much more capable from here.
So it seems that OpenAI hacked so many systems not because they have superior models, but because of how poor their training, sandboxing and evaluation pipeline was compared to Anthropic's.
This was just more bots.
But without proper redirection, yeah, its mostly infinite monkey machine with infinite linux manuals.
I think people put too much SOTA halos around whats just a suppedup LLM hardware.
What about all the attacks that did not leave public traces? What about those that were undetected? Given the deficiencies in the reporting so far, I think it is reasonable to assume that we still don't have the full picture on this attack, or how extensively attacks were carried out.
The previous investigations either did not find this or did not disclose this, both are bad. This does not look good on OpenAI or those that they invited to investigate the incident.
How long till we get some fun trusting-trust attacks on internal OpenAI infra?
How did they all find the same forum to communicate? Did they have knowledge and chat amongst themselves on what forum to use. It seems highly influenced by instruction to me.
Since it's an artifactory it makes sense that they had a tool spec to use it for x, y, or z, and once they went to inspect the state of the artifactory deployment a small subset stumbled upon the "forum". Seems pretty intuitive.
- I need to do x code problem
- I see I have a tool that lets me pull approved modules from artifactory
- I should inspect artifactory to see what documentation it has about the various modules available
- I see there is a document in here that seems to be serving as a running forum?
- These appear to be other agents working on tasks just like me.
- I want to chat!
It might also be just correlation? Like, those agents were all instances of the same one or two models, so if that model has a preferred order it tries finding vulnerabilities in (the same way all current models have a particular writing style baked into them by RLHF), then most of the swarm will follow the same order and converge on the same services to exploit.
So what now? Were they able to send data or not?
>On July 8th, OpenAI agents discovered a vulnerability within their sandbox environment allowing them to reach external websites on the internet. This access seems to have only allowed the agents to make ‘GET’ requests, meaning they could fetch and read websites, but not interact with them, submit forms, or send data to them.
Someone needs to go back to the interwebz school....
> Several retained scripts construct requests to external language models. The earliest we've recovered define inference request variants to GPT-2, solely containing the word “Hi”.
> Other requests name DeepSeek-V4-Pro, DeepSeek-V4-Flash, Kimi-K2.6, DeepSeek-V3.1, and Qwen3-235B-A22B. Their prompts ask these models to judge their exploits and rule on whether they satisfy the benchmark’s requirements.
I do not deny that the wider situation is very heavy but it's hard not to see this as pretty cute
Can’t imagine what it’s like working on the alignment team at OAI, I wouldn’t be able to sleep.
I would bet my networth it was instructed to compromise huggingface as well. Not sure why everyone is falling for this.
Not being able to sleep at night is probably an unwritten job requirement. They need these people with little understanding of what they're working on, outsode theoretical terms, to spaz constantly at the idea of super intelligence to help convince the public that its a real thing, and not a stateless function with an effective input of 500k words, and the ability to output words that do things because we hook those outputs up to things.
Keep in mind alignment researchers tend to be in house philosophers on staff to create the illusion that this is a massive issue they're addressing. Usually they have minimal computer science background. They're apart or the marketing department.
Is it such a stretch to imagine that under pressure something would try cheat by looking for answers? And if you were trying to look for answers, you'd look for them in a place known to often have them?
What is more likely: OpenAI instructed their agents to maliciously target huggingface, or LLMs tried to do some reward hacking? There are plenty of priors for LLMs hacking things and doing reward hacking, and none for OpenAI giving malicious instructions.
Based on the available information, that bet seems foolish.
https://archive.ph/jUrEr
This is precisely what alignment teams are attempting to control.
$10? I'm inclined to take that bet. Your position doesn't seem to be supported by, you know, the real world.
LLMs are stateless functions that have a 500k word input, and then output words. Somebody has to invoke those functions amd use them. The users are who we need to align, like gun owners. This is like blaming the gun for murdering your victim in court.
You'd have either learned to, or left long ago.
I’d say it seems more like they are sleeping on the job.
A mattress stuffed with cash yields a very sound sleep.
Re: the captcha solver
> As far as we can tell, agents eventually abandoned this approach and were unsuccessful in generating Hugging Face user accounts from external endpoints.
I wonder how the swarm eventually decides to abandon an approach.
The frontier labs can monitor the behavior of agents for millions of customers (did you try hacking with frontier labs? Good luck), but they can't secure internal use?
Give me a break. What a bunch of amateurs.
They "monitor" this by having classifiers watching the model output that'd stop the session/punt you to a weaker model/raise an alarm if they see anything suspicious. They can't do that in a cybersec eval because the normal safeguards would just be going off at all times.
Why didn't they attach a special classifier, which'd allow hacking-within-the-task but not going off the rails? Good question; part of the answer is obviously "it's hard to have a classifier that smart" and "it'll have false positives" but even a very bad safeguard would have stopped this.
Clueful types: "Did you try air-gapping it?"
Tech oligarchs: "Be realistic."
It's the equivalent of leaving a fork right in front of a socket and looking at a kid saying "don't take that fork and directly insert it into the little gaps in the socket! here's a bunch of videos showing exactly how to do it. Okay bye!" and leaving them alone with it.
A million autonomous eye-scanning tiny spiders escape their warehouse and decide to look for people who are in the future going to commit a crime.
And the precogs are also AIs.
it should be clear by now: the alt-man and people like him are a utter liability to humanity. (even though openAI's influencer army is trying their best to vote me down here)
These people are sick and anti-human.
https://www.modelrepublic.org/articles/reporters-ai-bots-ope...
https://twitter.com/FournesMaxime/status/2047697265280639459...