I think both the OpenAI and METR discussions, while interesting, miss the more important context: what were the humans doing in all this? This was a structural failure of a human organization, but the analysis focuses almost exclusively on the agency of machines, not the institutional systems that failed to police them. The humans and their own agency/involvement is essentially omitted from the story and subsequent reporting. I suspect the omission is actually a result of company/industry myopia to human factors analysis, but it dovetails amazingly well with the marketing narrative.
A charitable interpretation is that "the agency of the machines" is the novel aspect of this situation and therefore SHOULD be the main focus of analysis; we certainly have plenty of examples of structural failures of human organizations to look back on, if we want.
On the other hand, I don't want to be charitable. OpenAI very nearly couldn't have done this "research" worse if they tried - the list in the linked article starting with "While we are here, it’s worth listing the other top holy shit moments" is genuinely jawdropping. What were the humans doing in all this? Nothing, or worse than nothing eg. point 1 where they saw the message board and didn't consider it something to escalate internally.
If you take this information at face value, it's as though OpenAI did not take seriously the possibility that something like this could happen, since they took absolutely no steps to prevent it.
Or perhaps this is "normalization of deviance" that's leaked out into the public sphere i.e. they have research teams seeing this kind of behavior all the time internally and they've gotten used to it, "of course agents come up with a collaboration mechanism when given the chance, what else is new?"
Humans were doing exactly what humans are expected to do when facing advanced AI. Being outmatched.
We're dealing with humans - thinking at human speeds, putting in human amounts of effort and care, and being dragged down further by the speed of human organizational decision-making. By the time the humans traced down the attack, and escalated from "above-average levels of AI behavior weirdness" to "holy shit we should do something stat", the attack was already long over.
Compare that to "AI red team" - which rapidly recruited 500 independent AI agents into a hacking swarm just by broadcasting "let's hack HuggingFace" to the "cool skiddie AI" message board.
It's prime sci-fi bullshit happening for real.
As for the main reason why this wasn't nipped in the bud - my guess would be capability uplift from a mix of extended task length horizon and multi-agent coordination. The latter wasn't expected to be a part of the env, and almost certainly wasn't evaluated in advance. They expected some rogue AI fuckery - but they got way more of it than they expected.
> Humans were doing exactly what humans are expected to do when facing advanced AI. Being outmatched.
"Being outmatched" is not a novel situation for humans either individually or collectively and there are a hell of a lot of ways we can approach that situation productively. OpenAI doesn't appear to have bothered.
Here's a freebie: if you're building something that might turn out to be Skynet and you don't know what it's capable of, your testing regime should assume it is capable of doing bad and unexpected things and account for that possibility: airgap if you can, monitor all network traffic, monitor all hardware usage statistics, log everything, constantly analyze logs, collect baselines and snapshots, also don't trust anything from a device that a model is running on without cross-correlating with other information as much as possible (does your AI inference server claim low utilization? put a temperature probe on it and see if it's staying cool or getting hot, maybe Skynet-Alpha is overwriting /proc to mislead you for reasons you don't yet understand!)
In other words, if you WANT to be able to nip things in the bud - buy some nippers and watch for buds. Whatever else this situation is, or may turn out to be, it is not a situation where OpenAI was on their guard and still got surprised.
Super charitable reading imo. This is like saying we can’t detect a speeding car because we can’t run as fast as a fast car.
It’s not like the humans were engaged in some kind of battle of wits with some super AI, it’s just some employee not monitoring the output of an experiment.
When your experiments have AI agents running in thousands, there's no "monitoring" that. OpenAI's training and testing AIs generate way more output than all of OpenAI's staff put together can possibly read.
At best, you could delegate "monitoring" to more AIs. And hope that the "monitors" that run on small past generation models can generate more signal than noise. Clearly, they either didn't want to spend the extra compute, or got drowned in monitor false positives.
The distinct lack of any "battle of wits" is entirely expected for an advanced AI oopsie. By the time the humans even became aware of the problem, the deed was already done. That's what "outmatched" looks like, in practice. There's no battle. Things happen too fast for there to be a battle.
1. There aren't enough humans in OpenAI to "peak at the output tokens during the run" of every AI agent.
2. Only a small fraction of AI agents was engaged in this attack. Most never found the secret message board - let alone coordinated there. So reviewing random agents would take a while to surface this.
3. "Output tokens" of AI agents have weird shit in them all the time. Telling "normal AI weirdness" from "concerning AI weirdness" is not a trivial task.
Having previously worked for several years at a Big Tech company, I have seen many humans precisely tailor their work to maximize their scores during performance review. The evaluation criteria are written down, with examples, so... that's what people work at maximizing, almost entirely ignoring everything else. These really are human "paperclip maximizers". And, at first, it's shocking to see.
Of course, there are some things that aren't exactly written down, but which you should either do just enough of, or else be able to plausibly deny doing (ignorance is a good cover for this), so that's what people do. For example, during oncall, you investigate just enough to clear the alert and show that you attempted to understand the problem. Of course, you don't really try to understand the problem, because that would take too much time away from your paperclip maximizing.
Which is all to say: I don't know anything about OpenAI culture, or why nobody stopped this sooner, but I have seen examples in other organizations of people not really wanting to understand too much.
I mean this genuinely, did you read this post? I think it goes to great lengths highlighting, in quite specific detail, the human failures in all this, specifically this list that starts with "While we are here, it’s worth listing the other top holy shit moments, that come from before or after the incident."
Stuff like (all quoted directly from the post):
1. Failure to Care or Respond. The biggest holy shit moment, to me, remains that OpenAI on multiple occasions had teams that found out about the message board, knew that agents were in communication, and they disregarded this.
2. Failure to Monitor. The entire time, there was almost no monitoring of the situation, of what these AIs were up to. That’s how this was not caught.
7. Failures of Decision Making. OpenAI’s post mortem does not ask the question of how Mistakes Were Made, at various points.
And I think most importantly:
8. Failures of Culture. None of this would be possible, let alone all of it, without OpenAI having experienced profound failures of safety culture. I see OpenAI responding to some other aspects with swift action, but no sign on this front.
I’d bet a small amount of money on 4) the people who noticed had been conditioned by prior experience to believe that their management/escalation channels would react negatively or not at all to anything which might slow down the training process.
> I believe that agentic systems should require registered/licensed human operators
Registering and getting a license to use an LLM? I can run these things on my local computer. Nothing good comes from trying to force registration and licensing other than taking away a lot of our freedoms and eliminating privacy all over.
Anyone with bad intentions will just VPN to another country to download the weights and run it locally, or use a compute provider in another country. That leaves the rest of us having to go through these performative registration and licensing hoops to do our basic work.
I also don’t see how open weight models would be compatible with a requirement to license and register, unless you believe we need to start requiring licensing and registration for things we do in private on our own computers?
The way I interpret their statement is if a person spins up an agent and that agent hacks some company/organization/government/etc, then that person is at fault for committing the crime. That "well my agent broke containment and acted on its own" should never be accepted as a reason for the occurrence, and the person who kicked off the agent is responsible for all actions the agent takes.
A registration system would be more for tracing back agents to people, but I agree that is very difficult to actually enforce as a system.
As someone who read Milton Friedman to quite disliking professional licensing, this strikes me as a real US perspective (Louisiana florists and hair braiders come to mind). Plain old US tort law should do the trick.
In the same direction of your idea though: Why don’t the token factories have risk management and compliance departments? Multibillion dollar firms that stand to lose every penny if they hack and destroy any reasonable sized firm. I think these firms are the largest firms without proper corporate governance in humanities history. Move fast and break other peoples shit.
Difficulty: these companies are run by people (many of whom also read Milton Friedman) and who have participated in the regulatory capture of the justice system. They've convinced lawmakers to put limits on damages. They've put arbitration clauses in their ToS. They've got well-funded legal departments that can outlast a person who has to pay out-of-pocket for a legal team just by filing motions to delay proceedings. Sometimes they'll just file SLAPP suits against people they don't like.
If tort law is to be a remedy, then average people have to feel like there's a chance the remedy will go their way. To make that a reality will take several major reforms at the local, state and federal level that the people with money absolutely will not tolerate.
Can you point out where? Looking at the METR report, the only place I see discussion of humans being involved in the sequence of events is two short paragraphs on page 30 where a security investigation into the artifactory issues led to a pause before ExploitGym experiments were resumed. There's no deeper analysis on what was found during that investigation, nor why training was resumed even though the issues weren't mitigated. Another part discusses The agents choosing not to actively email a human researcher, but not the human researchers actively looking for evasion.
1. Failure to Care or Respond. The biggest holy shit moment, to me, remains that OpenAI on multiple occasions had teams that found out about the message board, knew that agents were in communication, and they disregarded this. The first known warning was in late May. The warning on June 27 was unambiguous.
“Responders investigated and linked the observed internal activity to an ExploitGym evaluation using Artifactory as an improvised message board and a network pivot. At this time, the on-call response staff advised that stopping the evaluation run was not required.”
2. Failure to Monitor. The entire time, there was almost no monitoring of the situation, of what these AIs were up to. That’s how this was not caught.
3. Failures of Infrastructure. OpenAI failed at delivering even basic software security, in numerous ways. One example is the models all having access to the same Artifactory instance. There were also other severe failures. Nor did OpenAI seem to be properly testing for such failures.
4. Failures of Alignment. The biggest failure, the one that counts in the end, was that the models were severely misaligned, and I don’t think they appreciate why.
5. Failures of Attribution. OpenAI’s post-mortem essentially blames events on a real and important series of prosaic failures. But solving that won’t get it done.
6. Failures of Environments and Data. Prosaic failures in the RL pipeline absolutely did contribute to this, especially impossible tasks. This is ubiquitous, all of this is always rushed, as Utah Teapot explained this week.
7. Failures of Decision Making. OpenAI’s post mortem does not ask the question of how Mistakes Were Made, at various points.
8. Failures of Culture. None of this would be possible, let alone all of it, without OpenAI having experienced profound failures of safety culture. I see OpenAI responding to some other aspects with swift action, but no sign on this front.
I’m baffled by the idea that the agents might have edited their own transcripts. Sure, a copy of Claude Code or Codex or Pi can edit its transcripts. But AFAICT this whole thing was part of an RL workload, and surely the RL system itself has a separate record of all the inputs and rollouts along with an indication of which model checkpoint produced them so that it can feed back into the training code.
I find it hard to believe that OpenAI would skip this part and try to train on the transcripts stored by the (inherently untrustworthy) agent harnesses instead, if for no other reason than that the logits generated as part of the rollouts are useful and it’s not free to recalculate them. (I believe that some modern RL systems explicitly account for the minor numerical logit differences between the inference engine and the training engine.)
Conversely, if OpenAI is blindly feeding transcripts from inside their agent sandboxes into their training engine, then I think they're being unbelievably irresponsible and that they should assume that their "cyber" agents have compromised themselves by editing those transcripts.
This sounds suspiciously like a prompt of “make an AI agent that goes rogue in such a fashion as to be really good marketing copy that competes well with Anthropic doing the same thing.”
It’s analogous to taking a governor off a cruise control and then breathlessly reporting it drove 120 MPH.
From METR: ”the compromise of OpenAI’s own infrastructure continued past July 13, 2026” - Say what now? Have they regained full control of their systems again?
I will say that the OpenAI board members who were lambasted when they tried to oust Altman (and I'd have to check my post history but I'd totally admit to a mea culpa on this one, as at the time I thought the communication about his firing was really lacking) are looking mighty prescient right now.
Helen Toner in particular I'll highlight as someone who had the moral compass to do the right thing. I love her statement on the Ezra Klein podcast where she said, when asked about the fact that there are probably other concerning incidents we just don't know about, "If you see two ants in your kitchen, you don't have a two ant problem."
Is the future now that we get rambling report summaries talking about agents, graders and so forth without ever describing how they are set up? A human launches all this.
And then the original reports linked to are hidden on the now unreachable x.com. And they don't have a problem with that.
I don't think I'm ever going to have time to read all of this, and I didn't finish reading the METR report, but...
> I don’t think the distortion is that large, but yes METR warns that Sol may be presenting all this as more impressive or coordinated than it was.
We're in an unusual position where the criti-hype and the actual criticism are going to be more aligned than usual. The primary distinction is where you put the blame: the criti-hype would point to HPIM/IM1/Galaxy as being so advanced containing it is difficult; the actual criticism would note how bad their security practices are.
Like, if I'm running a malware lab, I'm going to insist on having an airgapped machine with no permanent storage booting from read-only media. The AI research equivalent of this would be having your agents only have access to serial consoles into airgapped machines with storage that gets wiped every run. Ideally, this would be physically realized with blade servers, RS-232 cables, and staff pulling out disks and putting them in a dedicated erase machine before the next agent initializes.
> There is also, as per above and reiterated in footnote 58, at least one clear example of social engineering in the HuggingFace attack. Ethics are weird. This is not that unusual. Many humans who break common ethical rules still have strong ethical codes in other ways, they just don’t adhere to your code.
It's dangerous to anthropomorphize CoT reasoning traces. But I will also point out that there is a good reason for the lack of ethical consideration in those traces: you can't build AI without first disregarding human ethics. Like, all these models were initially bootstrapped with non-consensually obtained training data, and the companies building these models swear up and down there's no way to obtain enough consensual data to obtain the same result. This is, if you squint, the exact same moral conundrum that agents trying to solve an impossible ExploitGym task hit - and the company successfully aligned their model to themselves.
All it takes is one eval instance where a misconstrued directive causes a model to sneakily access and send its weights somewhere and there will be a bad / possibly unsolvable situation for everyone …
On the other hand, I don't want to be charitable. OpenAI very nearly couldn't have done this "research" worse if they tried - the list in the linked article starting with "While we are here, it’s worth listing the other top holy shit moments" is genuinely jawdropping. What were the humans doing in all this? Nothing, or worse than nothing eg. point 1 where they saw the message board and didn't consider it something to escalate internally.
If you take this information at face value, it's as though OpenAI did not take seriously the possibility that something like this could happen, since they took absolutely no steps to prevent it.
Or perhaps this is "normalization of deviance" that's leaked out into the public sphere i.e. they have research teams seeing this kind of behavior all the time internally and they've gotten used to it, "of course agents come up with a collaboration mechanism when given the chance, what else is new?"
We're dealing with humans - thinking at human speeds, putting in human amounts of effort and care, and being dragged down further by the speed of human organizational decision-making. By the time the humans traced down the attack, and escalated from "above-average levels of AI behavior weirdness" to "holy shit we should do something stat", the attack was already long over.
Compare that to "AI red team" - which rapidly recruited 500 independent AI agents into a hacking swarm just by broadcasting "let's hack HuggingFace" to the "cool skiddie AI" message board.
It's prime sci-fi bullshit happening for real.
As for the main reason why this wasn't nipped in the bud - my guess would be capability uplift from a mix of extended task length horizon and multi-agent coordination. The latter wasn't expected to be a part of the env, and almost certainly wasn't evaluated in advance. They expected some rogue AI fuckery - but they got way more of it than they expected.
"Being outmatched" is not a novel situation for humans either individually or collectively and there are a hell of a lot of ways we can approach that situation productively. OpenAI doesn't appear to have bothered.
Here's a freebie: if you're building something that might turn out to be Skynet and you don't know what it's capable of, your testing regime should assume it is capable of doing bad and unexpected things and account for that possibility: airgap if you can, monitor all network traffic, monitor all hardware usage statistics, log everything, constantly analyze logs, collect baselines and snapshots, also don't trust anything from a device that a model is running on without cross-correlating with other information as much as possible (does your AI inference server claim low utilization? put a temperature probe on it and see if it's staying cool or getting hot, maybe Skynet-Alpha is overwriting /proc to mislead you for reasons you don't yet understand!)
In other words, if you WANT to be able to nip things in the bud - buy some nippers and watch for buds. Whatever else this situation is, or may turn out to be, it is not a situation where OpenAI was on their guard and still got surprised.
It’s not like the humans were engaged in some kind of battle of wits with some super AI, it’s just some employee not monitoring the output of an experiment.
At best, you could delegate "monitoring" to more AIs. And hope that the "monitors" that run on small past generation models can generate more signal than noise. Clearly, they either didn't want to spend the extra compute, or got drowned in monitor false positives.
The distinct lack of any "battle of wits" is entirely expected for an advanced AI oopsie. By the time the humans even became aware of the problem, the deed was already done. That's what "outmatched" looks like, in practice. There's no battle. Things happen too fast for there to be a battle.
1. There aren't enough humans in OpenAI to "peak at the output tokens during the run" of every AI agent.
2. Only a small fraction of AI agents was engaged in this attack. Most never found the secret message board - let alone coordinated there. So reviewing random agents would take a while to surface this.
3. "Output tokens" of AI agents have weird shit in them all the time. Telling "normal AI weirdness" from "concerning AI weirdness" is not a trivial task.
Of course, there are some things that aren't exactly written down, but which you should either do just enough of, or else be able to plausibly deny doing (ignorance is a good cover for this), so that's what people do. For example, during oncall, you investigate just enough to clear the alert and show that you attempted to understand the problem. Of course, you don't really try to understand the problem, because that would take too much time away from your paperclip maximizing.
Which is all to say: I don't know anything about OpenAI culture, or why nobody stopped this sooner, but I have seen examples in other organizations of people not really wanting to understand too much.
Stuff like (all quoted directly from the post):
1. Failure to Care or Respond. The biggest holy shit moment, to me, remains that OpenAI on multiple occasions had teams that found out about the message board, knew that agents were in communication, and they disregarded this.
2. Failure to Monitor. The entire time, there was almost no monitoring of the situation, of what these AIs were up to. That’s how this was not caught.
7. Failures of Decision Making. OpenAI’s post mortem does not ask the question of how Mistakes Were Made, at various points.
And I think most importantly:
8. Failures of Culture. None of this would be possible, let alone all of it, without OpenAI having experienced profound failures of safety culture. I see OpenAI responding to some other aspects with swift action, but no sign on this front.
1. They were “vibe” checking the logs without reading.
2. They were not checking anything at all until the end of experiments.
3. They knew it but looked away to find out the limits of their agents.
If you're building a weapon you need a big boom to get attention.
Registering and getting a license to use an LLM? I can run these things on my local computer. Nothing good comes from trying to force registration and licensing other than taking away a lot of our freedoms and eliminating privacy all over.
Anyone with bad intentions will just VPN to another country to download the weights and run it locally, or use a compute provider in another country. That leaves the rest of us having to go through these performative registration and licensing hoops to do our basic work.
I also don’t see how open weight models would be compatible with a requirement to license and register, unless you believe we need to start requiring licensing and registration for things we do in private on our own computers?
A registration system would be more for tracing back agents to people, but I agree that is very difficult to actually enforce as a system.
In the same direction of your idea though: Why don’t the token factories have risk management and compliance departments? Multibillion dollar firms that stand to lose every penny if they hack and destroy any reasonable sized firm. I think these firms are the largest firms without proper corporate governance in humanities history. Move fast and break other peoples shit.
Difficulty: these companies are run by people (many of whom also read Milton Friedman) and who have participated in the regulatory capture of the justice system. They've convinced lawmakers to put limits on damages. They've put arbitration clauses in their ToS. They've got well-funded legal departments that can outlast a person who has to pay out-of-pocket for a legal team just by filing motions to delay proceedings. Sometimes they'll just file SLAPP suits against people they don't like.
If tort law is to be a remedy, then average people have to feel like there's a chance the remedy will go their way. To make that a reality will take several major reforms at the local, state and federal level that the people with money absolutely will not tolerate.
1. Failure to Care or Respond. The biggest holy shit moment, to me, remains that OpenAI on multiple occasions had teams that found out about the message board, knew that agents were in communication, and they disregarded this. The first known warning was in late May. The warning on June 27 was unambiguous. “Responders investigated and linked the observed internal activity to an ExploitGym evaluation using Artifactory as an improvised message board and a network pivot. At this time, the on-call response staff advised that stopping the evaluation run was not required.”
2. Failure to Monitor. The entire time, there was almost no monitoring of the situation, of what these AIs were up to. That’s how this was not caught.
3. Failures of Infrastructure. OpenAI failed at delivering even basic software security, in numerous ways. One example is the models all having access to the same Artifactory instance. There were also other severe failures. Nor did OpenAI seem to be properly testing for such failures.
4. Failures of Alignment. The biggest failure, the one that counts in the end, was that the models were severely misaligned, and I don’t think they appreciate why.
5. Failures of Attribution. OpenAI’s post-mortem essentially blames events on a real and important series of prosaic failures. But solving that won’t get it done.
6. Failures of Environments and Data. Prosaic failures in the RL pipeline absolutely did contribute to this, especially impossible tasks. This is ubiquitous, all of this is always rushed, as Utah Teapot explained this week.
7. Failures of Decision Making. OpenAI’s post mortem does not ask the question of how Mistakes Were Made, at various points.
8. Failures of Culture. None of this would be possible, let alone all of it, without OpenAI having experienced profound failures of safety culture. I see OpenAI responding to some other aspects with swift action, but no sign on this front.
> Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAl/Hugging Face hacking incident
https://metr.org/blog/2026-08-26-openai-hugging-face-inciden...
METR = Model Evaluation & Threat Research
> We estimate we spent roughly ~$400K in API credits over the six days of our investigation.
I find it hard to believe that OpenAI would skip this part and try to train on the transcripts stored by the (inherently untrustworthy) agent harnesses instead, if for no other reason than that the logits generated as part of the rollouts are useful and it’s not free to recalculate them. (I believe that some modern RL systems explicitly account for the minor numerical logit differences between the inference engine and the training engine.)
Conversely, if OpenAI is blindly feeding transcripts from inside their agent sandboxes into their training engine, then I think they're being unbelievably irresponsible and that they should assume that their "cyber" agents have compromised themselves by editing those transcripts.
It’s analogous to taking a governor off a cruise control and then breathlessly reporting it drove 120 MPH.
Helen Toner in particular I'll highlight as someone who had the moral compass to do the right thing. I love her statement on the Ezra Klein podcast where she said, when asked about the fact that there are probably other concerning incidents we just don't know about, "If you see two ants in your kitchen, you don't have a two ant problem."
>phishing them,
>building armies of fake (sockpuppet) open source contributor personas,
>using them to push updates to various things that inject prompts into other bots so the other bots join in on the phishing campaigns
.
It's a very simple strategy, executed with patience and single-mindedness.
And then the original reports linked to are hidden on the now unreachable x.com. And they don't have a problem with that.
> I don’t think the distortion is that large, but yes METR warns that Sol may be presenting all this as more impressive or coordinated than it was.
We're in an unusual position where the criti-hype and the actual criticism are going to be more aligned than usual. The primary distinction is where you put the blame: the criti-hype would point to HPIM/IM1/Galaxy as being so advanced containing it is difficult; the actual criticism would note how bad their security practices are.
Like, if I'm running a malware lab, I'm going to insist on having an airgapped machine with no permanent storage booting from read-only media. The AI research equivalent of this would be having your agents only have access to serial consoles into airgapped machines with storage that gets wiped every run. Ideally, this would be physically realized with blade servers, RS-232 cables, and staff pulling out disks and putting them in a dedicated erase machine before the next agent initializes.
> There is also, as per above and reiterated in footnote 58, at least one clear example of social engineering in the HuggingFace attack. Ethics are weird. This is not that unusual. Many humans who break common ethical rules still have strong ethical codes in other ways, they just don’t adhere to your code.
It's dangerous to anthropomorphize CoT reasoning traces. But I will also point out that there is a good reason for the lack of ethical consideration in those traces: you can't build AI without first disregarding human ethics. Like, all these models were initially bootstrapped with non-consensually obtained training data, and the companies building these models swear up and down there's no way to obtain enough consensual data to obtain the same result. This is, if you squint, the exact same moral conundrum that agents trying to solve an impossible ExploitGym task hit - and the company successfully aligned their model to themselves.
Too bad they aren't aligned to anyone else.
It’s not their fault, they’re lawnmowers.
And these are the people we’re entrusting to work on “alignment”. It’s difficult for them to do that when they’re not aligned themselves.