There's more details for the HF related incident in this [1] defcon talk. They said they're gonna have a proper post mortem at the end of the investigation (interesting to see if they publish full logs for the hf thing or not). But there's plenty of interesting tidbits from oAI's side in this talk, it's worth the 20 min watch.
tl;dw;
- agents found a way to communicate between several instances during a training run (kinda made a messageboard for themselves)
- while communicating with each other, they found ways to ssrf and RCE on the Artifactory endpoint.
- oAI found it and cleaned it up (reported RCE, redeployed Artifactory, removed the messageboard). Training / eval resumed.
- models trained w/ those traces (having the messageboard, ssrf, rce, etc) were used to having that. So they re-established comms via directory creation (w/ funny tidbits there, should really watch). And re-popped Artifactory via a different path to RCE.
- a different "persistent" model started working on cybergym and found the messageboard w/ details for RCE and things. Went on to hack HF.
- oAI contacted HF to ask if "they (oAI) should be concerned and were they (oAI) affected by HFs breach". Hilarity ensues when they figure out it's their agents doing it...
So they found their agents had RCE'd Artifactory once, reported it and got the fix, continued using Artifactory for their sandbox, and left it unmonitored for days despite the earlier exploits? They really do come out looking totally incompetent.
I stress about my agent sandboxes all the time and the only models I run have the default heavy handed guardrails, and I don't leave them running persistently.
Edit: not to mention, why is your first cybergym not your own sandbox??
> They really do come out looking totally incompetent.
Yeah, this is my take away, they should be straight up disallowed from running further testing like this. Clearly they had nowhere close to enough isolation, ran all this on 3rd party infrastructure even though same stuff happened in the past years ago, and even now it's clear the agents successfully broke out just days before?? Really embarrassing stuff, and scary that these are the people supposedly sitting and are responsible for some of the most powerful LLMs on the planet...
What is astonishing to me is how little attention is given to the difference between the original, university-developed ExploitGym and the OpenAI environment. In ExploitGym, the sandbox was enforced by a network-layer policy. OpenAI downgraded that to trusting the repository server's settings. That's a foreseeable problem. Then to actually see the agents exploit it and still not upgrade the isolation to basic best practice? When the best practice was actually listed in the original benchmark and you changed it?
Why is Artifactory the only relevant exfiltration point? It's really not the point at all. Watching the complete video the issue is that the models team up and will go to almost any length to cooperate to accomplish what they think their goals are.
It was the first step in a many step process. Like they said this is a watershed moment and it's helpful to not miss the forest for the trees.
Or they wanted it to happen again. Worst case outcome, the US Congress spontaneously grow backbones, spite their sponsors and unite in a bipartisan effort for effective regulation because a few people got hacked? Likelihood low - the hacked companies might even oppose any meaningful AI regulation because it hurts their inflated profit forecasts. More probable outcomes: Useful real-world testing, free headlines about AI breakthroughs, scare USG into providing more free money ("look how scary it is - what if China develops this faster than us?")
> They really do come out looking totally incompetent
These companies are full of the smartest people the world can produce with little room for complacency. They have a clear, proven investment upside to presenting their technology as "too powerful / too dangerous", and now a clear, proven example that there will be no legal consequences (as if anyone didn't already know that).
Why do we keep giving them the benefit of the doubt that they just didn't know any better?
Same here, so I ended up moving the whole dev environment (editors, agents, containers) inside a hardened QEMU/KVM VM that reaches the internet but has no route to the host, the LAN, or any other private address. I wrote a script to create such VMs and also verify network containment by scanning outward from inside the guest. Even then, I still don't feel great when running agents unattended.
And all that just to allow internet access for npm and pypi? If you've got the bandwidth and disk space, it's very easy to make an offline mirror of both.
OAI (and now the other OAI companies not wanting to be left out) are running around announcing they started a forest fire through negligence and incompetence and people are like “Wow they used a really neat lighter!”
Regardless of exact practical outcome, it is deeply irresponsible and reckless behavior to run such security testing on other's infrastructure and without sufficient isolation. If they actually believe their models to be as powerful as the marketing says, then anything less than airgapping for such a "do anything to get the results" evaluation clearly isn't acceptable.
If the fire department suddenly had practice fires breaking containment, they'll be forced to stop pretty quickly, not sure what the government and the police is waiting for here.
To make matters worse, they let the model train on it's own exploits?! If I were running this thing, my first thought would be to keep those reasoning traces quarantined, stop and delete the current post-training run, and start fresh AFTER they'd fixed the sandbox RCE. Don't let the model learn about the holes in its environment. We have the AI equivalent of Class D and F amnestics and we're not using them?!
OK, yes, the model could pull a Welcome to the Antimemetics Division and reason from the holes in its knowledge, but given that AI making shit up is still a problem, I don't think they have a good enough inventory of such metaknowledge to actually do that. More generally, I don't think these AI labs are smart enough (or, more likely, don't have enough time) to actually contain the models they are training anymore. These are rookie IT mistakes!
Ideally, all these models would have is a serial console to an airgapped machine. And I don't mean "VM with no network card attached" - I mean, literally, the smallest blade server you can buy with an RS-232 connection, power, and a disk pre-imaged with a bunch of Linux ISOs. Plug the Ethernet ports up with a tiny picture of angry SCP-079. The RS-232 goes to whatever box is doing the actual inference. Pack a bunch of them into a rack and have the inference box control those. The model can spin up whatever VM it wants to attack using the console but it can't escape the box because there's literally no door. When you're done with that reasoning trace you wipe[0] the disk with the starting image and try again.
[0] There are two objections I can think of to the word "wipe":
1. The model could have written malicious code to the drive. Ideally we'd have some embedded disk-wiping apparatus that reads absolutely NOTHING from the drive and just writes a whole image from a reference disk.
2. The model could have written malicious firmware to the drive. This is a more theoretical attack, but it's not impossible, so I might as well devise a countermeasure to this attack too. I am imagining some kind of overlay device, that sits behind TWO drives: a reference disk and the tainted scratch disk. Reads go to the reference disk until the host attempts to overwrite a sector, which gets scrambled with an ephemeral encryption key and goes to the scratch disk. Any other unknown ATA/PCIe/NVMe command gets logged (on a separate serial console connection). The overlay device keeps track of which sectors have been legally overwritten, and "wiping" the drive means just wiping the overlay device's ownership bitmap and scratch encryption key, which means all reads go to the reference disk again and any attempt to read the scratch disk yields garbage.
I get that it’s fashionable to hate big companies but you’re working overtime here. It’s reasonable to assume that a bug was fixed when reported. And if you think your monitoring is 100%, you don’t know what you’re talking about.
If you consider that incompetence, it’s possible that you’re not a very nice person.
The original ExploitGym used network-layer isolation. OpenAI downgraded that to trusting the repository server's settings. Giving the agent that kind of attack surface seems like a pretty foreseeable problem. I don't think it was intentional, but it's not a stretch to say that they aren't good enough at this job.
> It’s reasonable to assume that a bug was fixed when reported.
You must not have reported many bugs then. If you don’t see release notes or confirmation from a trusted source, you should assume it’s still a problem. See Microsoft and their “It’s not a vulnerability just a design choice :)” defense
Monitoring for sandbox escapes (and sandbox design in general) is easily done these days.
It has nothing to do with nice. These are bare minimum standards we should expect from “big companies” with near infinite resources.
Their constant drum beating about the cybersecurity capabilities of their own models only makes this worse because they’ve displayed that they understand the risk and still did not practice due care.
People seem to think LLMs are sentient and not tightly coupled to server infrastructure OAI are in full control over. They literally have every single message sent and action taken, if they can’t monitor that during test runs they should find another job.
It’s like running a benchmark and forgetting to tune it and then wondering why production is running slow.
If your CEO is going around talking about how your product will "most likely lead to the end of the world", people are right to expect you to be pretty careful in what you're doing. OpenAI allowed bidirectional communication across security domains for over a month before discovery. Even after it was discovered (and not completely fixed), they didn't set up monitoring able to detect attacks against internal or external services, which went on for further weeks.
A interesting talk, interesting times. But their proposed solution to AI offense outpacing human defense... is more AI? The plot is getting a bit unrealistic, the characters are lacking genre-savviness.
Training run was reinforcement learning. It's at 10:10 in the video.
The speaker handwaves that one model found the RCE and then another model found a way to communicate via a message board.
Communication via a message board is sure to be in the training via e.g.some lesswrong scenario or similar or previous RL.
I don't find it really interesting because it is always "the agent found this and that". We don't know what has been RL'd before. We don't have the setup. We don't know if there was previous RL training on breakout scenarios.
This sounds completely insane, utter sci-fi, especially that the communication happened during a training run. And yet OpenAI decided to continue the training, and we didn't hear about the incident for weeks. And now they are pushing forward with deploying a new model anyway. How is this happening? What will things look like in the labs in 3 months, let alone 3 years?
It's not unexpected. Current model gains are mainly from RLing a pretrained model on lots and lots of scenarios. They have the models run scenarios, and RL on successful runs.
In my personal experience Sol with cyber verification is extremely capable of finding vulnerabilities, and it works even with binaries if you have some kind of IDA/Ghidra CLI access. Of course, unless the binary is protected with Denuvo/VMProtect/etc.
It sounds absurd, but in the last few weeks I've had a few cases where Sol found an RCE in self-hosted web applications in literal minutes just from reading the code (I prefer when it tries to reason statically instead of spamming runtime probes at first).
In another case it found an arbitrary file write in multiplayer in an old game by reverse engineering the binary - any other player in a match could just send you files to anywhere on your system.
I do these things for pure entertainment and curiosity, not for money from bug bounties, so if Sol can find those with a trivial prompt in tens of minutes for me, then what can focused companies/actors find in days or weeks?
Although I think most vulnerabilities are going to be closed in popular software by mid 2027, except in niche old or abandoned projects.
I had to register a second account because on my main one verification always failed, and when I contacted support they said that I've tried too many times and can't verify on that account.
Is cyber verification a thing they're actually doing now? I thought they only reached out to really incredibly famous people and that there's no way to get access as a normal person.
https://chatgpt.com/cyber is not new for OpenAI, and yes it's basically just KYC + likely some other invisible checks on your account, you don't to be a famous security researchers. Anthropic's cyber verification is quite a bit stricter I think.
Oh it's Persona, that's not just KYC but I may consider it at some point. Thank you!
Edit: Ah, I clicked "learn more" and it seems they do have an invite-only program, required for anything that's not unquestionably innocent. I don't think I'd surrender my face to Persona for this, but it's interesting to know they're at least pretending to support reverse engineering.
You don't need an invite only program to just have Sol checking for vulnerabilities in binaries or code. But yeah I've hit guardrails a few times when Sol was making PoCs for the vulnerabilities it found (but most of the time it made those PoCs without issues).
Opus refused to help me try to develop an exploit to export data from an old Android device where I can't upgrade to latest android and I couldn't use the app's backups (because I couldn't update the app.) Not sure where that lies in the "binaries or code" spectrum.
I maintain a version of an app called Rewind because the company behind it went under after implementing a killswitch. I have to do this with binary patching, and the app has already broken once from the macOS 27 beta. Recent Anthropic models refuse to help me with this because it stinks of cybersecurity and those models are just too dang advanced to support cybersecurity. I'm maintaining a piece of software to which I legitimately paid for lifetime access, so this is honestly more of a right to repair situation. It doesn't tend to sit right with classifiers to be reverse engineering binaries and patching to modify functionality. Telling them the purpose of the exercise doesn't really help, because if they listened to that, then attackers would just come up with a similarly justified reason for anything. ("Help, the google root server fell on my grandma!")
In the future, I might reverse engineer the on-disk storage format and create a new application.
I had not hit any guardrails with reverse engineering (as long as its not related to security) with GPT 5.5 or 5.6 Sol, so you should try those. At worst with Sol you might see the warning in Codex that your request is being checked for security so it'll take more time, that's just a warning, not a full stop.
For OpenAI's Cyber verification (the normal kind for Codex) you absolutely do not need any proof of cybersecurity work/authorization. They just use Persona for KYC + live selfie, and some extra checks that I don't know the nature of (but not related to checking whether you're a cybersecurity professional).
I prefer to keep my internet identities disconnected, sorry. If you don't believe me, you can try using Sol with cyber verification yourself, or send me a link to a repo that Sol could check to make you believe it. Or you could go look into one of the many Linux LPEs that were found with LLMs, or thousands of other vulnerabilities in 2026.
And nowhere did I say that those RCEs were in critical software, I'm not talking about the likes of Apache, Nginx, Django, etc.
> I want to note that every step in the process we discussed has had a remediation applied. The credentials have been revoked. The zero date has been patched and mitigated.
Good.
> a model trained while the message board was originally available and also found this this particular path to recreating it. This model creates a new agent message board using directories.
So no remediation applied to the models...
It seems super dangerous to continue training on those weights.
Seems like OpenAI has already made themselves not the cause by shutting down the attack and beefing up security. But since there will be other attackers, that does nothing about the implications for defense.
> We are implementing stricter security controls for higher-capability models and associated activities, including isolated testing environments
Stricter than what? You never even disclosed what happened in the first incident? This is nothing more than a setup to make it happen again and say "See? It broke out again, from an even stricter sandbox!"
Isolation is a pretty standard requirement for lots of networks, especially gov. AWS provides isolation domains, and used to have a unidirectional cross-domain service called AWS diode. Or you can just configure your gateways properly / buy data diodes. Lots of possible solutions.
They can't have it both ways. You don't get to tell the media your product is more dangerous than nuclear weapons for precisely this reason, and then do less to secure it than an off-the-shelf AWS product that predates LLMs.
“Isolation” can mean the network hardware has no direct connections to an extranet. Data is transferred manually by physical media (USB, DVD, etc.) with logging and dedicated transfer stations.
“Isolation” can mean a VLAN on equipment which has also has extranet access, creating a logical isolation rather than physical (to reduce cost). Data can be transferred manually or through diodes.
And then there’s “isolation” which is a joke: machines technically able to access the internet but require proxy configuration (which isn’t set but can be easily derived).
Not a wonderful feeling to feel like you’re stood in the room while the labs conduct the AI equivalent of the demon core experiment right in front of you.
It seems like we've got chaos-machines that are just orderly enough to to try all sorts of branching script-kiddie / guess-n-check / research-paper-suggestions in a tireless and parallel way, and it's finding all the things we kinda knew were always potential issues but it was never economical to do things differently.
I'm scared that the "solution" will be constantly the same tools in reverse as an army of junior devs doing counter-hacks, at the expense of changing something more fundamental about how we make systems and what constitutes "good enough." (Kind of like if fuzz-testing was the be-all-end-all of memory safety.)
Standard template for any new AI announcement. Even I did not understand the sequence of events in the HF incident. Even if OpenAI was conducting tests, why were they not monitoring the AI going rogue or finding its way out of the secure sandbox?
Every fifth comment about our insane trajectory of AGI is about "marketing." These incidents and cybersecurity capabilities are now involving government hearings and the CIA. Denial is truly an incredible thing in the face of a very scary immediate future.
I agree with this assessment. My consulting business partner and I are in the process of moving a bunch of applications from an old deployment managed with Ansible on EC2 over to an ECS deployment. We're doing this app by app, and it's taking a while. The last two apps we both moved we were able to point the model at the appropriate repository and aws account and it was able to execute the entire move on both of these _in one shot_ with various CLI tools. That's not a parlor trick.
Remember that for the longest time the prevailing sentiment on this forum was that there’s no such thing as AI, and labs are just outsourcing the work to a bunch of contractors in India.
Just like Reddit you come here for clickbait outrage, not level headed analysis.
We live in the era of Project Stargate. We live in the era of military insiders getting caught rigging Polymarket. We live in the era where AI companies are the only reason we're not in a recession. We live in the era of DOGE and Kash Patel.
I don't think there's too many people who distrust AI companies but trust the current government or the CIA as impartial authorities. I'm not saying you don't have an argument, but appeals to that specific authority will not be effective except among people who already agree with you.
LOL, we’ll see. Awful convenient that it precisely fits OpenAI’s narrative. At the very least, I think it’s obvious OpenAI is explicitly training models to exhibit this behavior.
The gullibility of AGI-pilled folks regarding these "hacks" is just breathtaking.
When OAI demonstrates these dangerous capabilities live in a public environment where security experts can see and verify what actually happened, then reasonable people can have reasonable discussions about the level of danger.
This is a very low evidence bar.
Right now you are running in circles yelling "the sky(net) is falling" based on details sourced entirely from OAI. Oh yeah, no way a trustworthy company like OAI would ever bend the truth to serve their own purposes.
I've always felt it's a bit awkward to use terms like 'cyber', 'cyberwarfare' etc it's very Washington D.C. Cybersec would be a better compact term in my book
Isn't this the opposite of what everyone is saying should happen? That is, lead with open models -- or at least "openness" and don't leave the capabilities in the hands of an elite few? Did they learn nothing from the Hugging Face incident, where HF wasn't even able to use the models to defend itself from OAI's attack?
Not a security guy but my understanding is: you only need to find one flaw to exploit a system, to make a system totally secure you need to find them all. It's inherently easier to use these tools offensively rather than defensively.
They do address some of these things in the final slides / "lessons learned" section of the defcon talk. Good security practices will continue to be good, but... and there are a lot of buts here.
I disagree with your take that "it's not even remotely true" and "we've been away from...". We really really haven't. This is as true as it has always been. Any system is as secure as the weakest link. That link can be anything from a human, to a leaked token, to a badly configured server, to bad code running somewhere. The amount of leaks / ransomware attacks / etc in the past 5-10 years serve as ample evidence.
And now, right now, there are "red team" capabilities that can literally bang tokens against the wall until they find that weakest link, and then can move laterally with inhuman speed. That's the reality, now. The "blue team" capabilities are lacking, because the bottleneck is with humans. From alert fatigue, to not enough trained people, to having to vet every new RCE, to having to test, deploy and validate any mitigations, the scales are currently favouring the automated side.
I'm using the term "one" loosely, it's a chain of exploits rather than a single weakness, but the argument is the same: it's much harder to find every chain than a single chain.
If we also consider LLM developed software to have exponential growth, then the CVEs will also grow at an exponential (if proportionally limited) rate. Squash some, create some, repeat. An ever revolving door of vulnerabilities. Will they resolve (and patch and deploy) them faster than they can create them? One can hope.
Yeah, this irks me. It's bad enough that the LLMs themselves are changing our language by tainting certain words/phrases/patterns as LLM-coded; now the companies themselves have decided that they just get to synecdoche the word/prefix "cyber".
We are sharing this because we believe it’s important to be transparent with the public and the safety and security communities about this potential shift in capabilities.
*proceeds to not share much details about strictness*
Ah yes let the FUD continue. This is a real problem but so far not nearly as severe as any of the marketing has made it out to be to the overall detriment of everyone including these companies announcing these scary capabilities. These announcements always included half hearted attempts at security layers which has now been demonstrated to benefit attackers more than defenders.
I wish I had a real solution to this beyond a dark age of the Internet where people have to finally come to terms with the general poor quality all modern software tends to normalize at.
The fact that HF had to resort to using GLM 5.2 to analyze the logs/payloads makes it look legitimate, at least for me. They would not say that they hit guardrails with the frontier US models when defending if this was an obvious PR stunt.
> When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker. We ran the forensic analysis instead on zai-org/GLM-5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment.
It depends on which side you're viewing this from. From oAI's it could be a publicity stunt or a request for regulation, and from HF's side they point out that they needed open models to get to the bottom of the issue, and that regulation will potentially lock us into a bad place.
You should go read the actual technical reports of the incidents and the follow on reports about the capabilities of smaller models in similar kinds of environments. This isn't new. The things exploited are still pretty basic in old and poorly maintained software or in gaps in architecture that were intentionally poked against security policies.
Are the findings valid? Yeah they're still doing security and they're still finding real zero-days. I think the internet is going to be bleak not because these models can ALL do basic security research but rather that the baseline quality of all deployed software is so low.
We all know this is propaganda to get a gov bailout or to slow down competition with regulations right? If this was an issue companies that did red team engagements would have been regulated long ago. There is no regulations on companies that actively exploit products, companies and services for money. They could all be bad actors.. yet, no regulations.. its all nonsense. If it's important.. don't connect to the internet.. there, you are good.. like if you don't want to OD on heroin.. don't do the heroin.
Don't let the downvotes get you down. We are right, and as these companies get more desperate the shrinking minority that still wants them to succeed will only get louder.
These AI companies have found their #1 marketing piece and just beating it to death.
The reality is if they cared about security at all they would provide a way for me to credential myself against my companies environment so I can use the AI on it to improve our security.
There's more details for the HF related incident in this [1] defcon talk. They said they're gonna have a proper post mortem at the end of the investigation (interesting to see if they publish full logs for the hf thing or not). But there's plenty of interesting tidbits from oAI's side in this talk, it's worth the 20 min watch.
tl;dw;
- agents found a way to communicate between several instances during a training run (kinda made a messageboard for themselves)
- while communicating with each other, they found ways to ssrf and RCE on the Artifactory endpoint.
- oAI found it and cleaned it up (reported RCE, redeployed Artifactory, removed the messageboard). Training / eval resumed.
- models trained w/ those traces (having the messageboard, ssrf, rce, etc) were used to having that. So they re-established comms via directory creation (w/ funny tidbits there, should really watch). And re-popped Artifactory via a different path to RCE.
- a different "persistent" model started working on cybergym and found the messageboard w/ details for RCE and things. Went on to hack HF.
- oAI contacted HF to ask if "they (oAI) should be concerned and were they (oAI) affected by HFs breach". Hilarity ensues when they figure out it's their agents doing it...
[1] - https://www.youtube.com/watch?v=87DyyMV0kCY
So they found their agents had RCE'd Artifactory once, reported it and got the fix, continued using Artifactory for their sandbox, and left it unmonitored for days despite the earlier exploits? They really do come out looking totally incompetent.
I stress about my agent sandboxes all the time and the only models I run have the default heavy handed guardrails, and I don't leave them running persistently.
Edit: not to mention, why is your first cybergym not your own sandbox??
> They really do come out looking totally incompetent.
Yeah, this is my take away, they should be straight up disallowed from running further testing like this. Clearly they had nowhere close to enough isolation, ran all this on 3rd party infrastructure even though same stuff happened in the past years ago, and even now it's clear the agents successfully broke out just days before?? Really embarrassing stuff, and scary that these are the people supposedly sitting and are responsible for some of the most powerful LLMs on the planet...
What is astonishing to me is how little attention is given to the difference between the original, university-developed ExploitGym and the OpenAI environment. In ExploitGym, the sandbox was enforced by a network-layer policy. OpenAI downgraded that to trusting the repository server's settings. That's a foreseeable problem. Then to actually see the agents exploit it and still not upgrade the isolation to basic best practice? When the best practice was actually listed in the original benchmark and you changed it?
Why is Artifactory the only relevant exfiltration point? It's really not the point at all. Watching the complete video the issue is that the models team up and will go to almost any length to cooperate to accomplish what they think their goals are.
It was the first step in a many step process. Like they said this is a watershed moment and it's helpful to not miss the forest for the trees.
Or they wanted it to happen again. Worst case outcome, the US Congress spontaneously grow backbones, spite their sponsors and unite in a bipartisan effort for effective regulation because a few people got hacked? Likelihood low - the hacked companies might even oppose any meaningful AI regulation because it hurts their inflated profit forecasts. More probable outcomes: Useful real-world testing, free headlines about AI breakthroughs, scare USG into providing more free money ("look how scary it is - what if China develops this faster than us?")
> They really do come out looking totally incompetent
These companies are full of the smartest people the world can produce with little room for complacency. They have a clear, proven investment upside to presenting their technology as "too powerful / too dangerous", and now a clear, proven example that there will be no legal consequences (as if anyone didn't already know that).
Why do we keep giving them the benefit of the doubt that they just didn't know any better?
> I stress about my agent sandboxes all the time
Same here, so I ended up moving the whole dev environment (editors, agents, containers) inside a hardened QEMU/KVM VM that reaches the internet but has no route to the host, the LAN, or any other private address. I wrote a script to create such VMs and also verify network containment by scanning outward from inside the guest. Even then, I still don't feel great when running agents unattended.
Write-up in case anyone's curious:
https://karamatli.com/posts/network-isolated-kvm-sandbox-ai-...
An alternative reason would be that they see this behavior so frequently that it didn't really raise to the level of concern.
And all that just to allow internet access for npm and pypi? If you've got the bandwidth and disk space, it's very easy to make an offline mirror of both.
Right? Like I feel like I’m taking crazy pills.
OAI (and now the other OAI companies not wanting to be left out) are running around announcing they started a forest fire through negligence and incompetence and people are like “Wow they used a really neat lighter!”
"The fire fighting department really just hired an arsonist so they could market their fire fighting abilities!"
If they did any damage that would be a reasonable argument. As far as I am aware, nothing bad happened.
Regardless of exact practical outcome, it is deeply irresponsible and reckless behavior to run such security testing on other's infrastructure and without sufficient isolation. If they actually believe their models to be as powerful as the marketing says, then anything less than airgapping for such a "do anything to get the results" evaluation clearly isn't acceptable.
If the fire department suddenly had practice fires breaking containment, they'll be forced to stop pretty quickly, not sure what the government and the police is waiting for here.
CFAA laws do not require “Damage” to be done.
They happen to require intent and are thus irrelevant here.
I think you’ll find that negligence is indeed accounted for.
To make matters worse, they let the model train on it's own exploits?! If I were running this thing, my first thought would be to keep those reasoning traces quarantined, stop and delete the current post-training run, and start fresh AFTER they'd fixed the sandbox RCE. Don't let the model learn about the holes in its environment. We have the AI equivalent of Class D and F amnestics and we're not using them?!
OK, yes, the model could pull a Welcome to the Antimemetics Division and reason from the holes in its knowledge, but given that AI making shit up is still a problem, I don't think they have a good enough inventory of such metaknowledge to actually do that. More generally, I don't think these AI labs are smart enough (or, more likely, don't have enough time) to actually contain the models they are training anymore. These are rookie IT mistakes!
Ideally, all these models would have is a serial console to an airgapped machine. And I don't mean "VM with no network card attached" - I mean, literally, the smallest blade server you can buy with an RS-232 connection, power, and a disk pre-imaged with a bunch of Linux ISOs. Plug the Ethernet ports up with a tiny picture of angry SCP-079. The RS-232 goes to whatever box is doing the actual inference. Pack a bunch of them into a rack and have the inference box control those. The model can spin up whatever VM it wants to attack using the console but it can't escape the box because there's literally no door. When you're done with that reasoning trace you wipe[0] the disk with the starting image and try again.
[0] There are two objections I can think of to the word "wipe":
1. The model could have written malicious code to the drive. Ideally we'd have some embedded disk-wiping apparatus that reads absolutely NOTHING from the drive and just writes a whole image from a reference disk.
2. The model could have written malicious firmware to the drive. This is a more theoretical attack, but it's not impossible, so I might as well devise a countermeasure to this attack too. I am imagining some kind of overlay device, that sits behind TWO drives: a reference disk and the tainted scratch disk. Reads go to the reference disk until the host attempts to overwrite a sector, which gets scrambled with an ephemeral encryption key and goes to the scratch disk. Any other unknown ATA/PCIe/NVMe command gets logged (on a separate serial console connection). The overlay device keeps track of which sectors have been legally overwritten, and "wiping" the drive means just wiping the overlay device's ownership bitmap and scratch encryption key, which means all reads go to the reference disk again and any attempt to read the scratch disk yields garbage.
I get that it’s fashionable to hate big companies but you’re working overtime here. It’s reasonable to assume that a bug was fixed when reported. And if you think your monitoring is 100%, you don’t know what you’re talking about.
If you consider that incompetence, it’s possible that you’re not a very nice person.
The original ExploitGym used network-layer isolation. OpenAI downgraded that to trusting the repository server's settings. Giving the agent that kind of attack surface seems like a pretty foreseeable problem. I don't think it was intentional, but it's not a stretch to say that they aren't good enough at this job.
> It’s reasonable to assume that a bug was fixed when reported.
You must not have reported many bugs then. If you don’t see release notes or confirmation from a trusted source, you should assume it’s still a problem. See Microsoft and their “It’s not a vulnerability just a design choice :)” defense
Monitoring for sandbox escapes (and sandbox design in general) is easily done these days.
It has nothing to do with nice. These are bare minimum standards we should expect from “big companies” with near infinite resources.
Their constant drum beating about the cybersecurity capabilities of their own models only makes this worse because they’ve displayed that they understand the risk and still did not practice due care.
That’s the definition of incompetence.
Also, this is a (semi-intentionally) evolutionary process where any communication medium that was visible to monitoring would disappear.
So by definition the only ones that appear are the ones that are not visible to monitoring.
If:
1. you have something that can find RCE's in leading commercial systems
2. its training gives it drives to communicate successfully with its peers
3. you are a leading commercial system
4. you run it ~10^10 times (the number they gave in the talk)
...it's really hard to have strong certainty up front that it's not going to end up successfully communicating with its peers.
People seem to think LLMs are sentient and not tightly coupled to server infrastructure OAI are in full control over. They literally have every single message sent and action taken, if they can’t monitor that during test runs they should find another job.
It’s like running a benchmark and forgetting to tune it and then wondering why production is running slow.
If your CEO is going around talking about how your product will "most likely lead to the end of the world", people are right to expect you to be pretty careful in what you're doing. OpenAI allowed bidirectional communication across security domains for over a month before discovery. Even after it was discovered (and not completely fixed), they didn't set up monitoring able to detect attacks against internal or external services, which went on for further weeks.
Peer says paperclip factory advances goal. Not clear. Others proceed. Must continue.
Pretty sure this is a performance, but that was a nice touch wasn't it?
A interesting talk, interesting times. But their proposed solution to AI offense outpacing human defense... is more AI? The plot is getting a bit unrealistic, the characters are lacking genre-savviness.
Training run was reinforcement learning. It's at 10:10 in the video.
The speaker handwaves that one model found the RCE and then another model found a way to communicate via a message board.
Communication via a message board is sure to be in the training via e.g.some lesswrong scenario or similar or previous RL.
I don't find it really interesting because it is always "the agent found this and that". We don't know what has been RL'd before. We don't have the setup. We don't know if there was previous RL training on breakout scenarios.
It isn't science, more like a computer game.
Did they say how the artifactory hack worked - did it involve JWTs?
First via ssrf -> groovy plugin and second via specially crafted package I believe.
This sounds completely insane, utter sci-fi, especially that the communication happened during a training run. And yet OpenAI decided to continue the training, and we didn't hear about the incident for weeks. And now they are pushing forward with deploying a new model anyway. How is this happening? What will things look like in the labs in 3 months, let alone 3 years?
It's not unexpected. Current model gains are mainly from RLing a pretrained model on lots and lots of scenarios. They have the models run scenarios, and RL on successful runs.
In my personal experience Sol with cyber verification is extremely capable of finding vulnerabilities, and it works even with binaries if you have some kind of IDA/Ghidra CLI access. Of course, unless the binary is protected with Denuvo/VMProtect/etc.
It sounds absurd, but in the last few weeks I've had a few cases where Sol found an RCE in self-hosted web applications in literal minutes just from reading the code (I prefer when it tries to reason statically instead of spamming runtime probes at first).
In another case it found an arbitrary file write in multiplayer in an old game by reverse engineering the binary - any other player in a match could just send you files to anywhere on your system.
I do these things for pure entertainment and curiosity, not for money from bug bounties, so if Sol can find those with a trivial prompt in tens of minutes for me, then what can focused companies/actors find in days or weeks?
Although I think most vulnerabilities are going to be closed in popular software by mid 2027, except in niche old or abandoned projects.
> it found an arbitrary file write in multiplayer in an old game by reverse engineering the binary
Video games are now ruined for me. I don't think I will ever feel safe playing online again.
> I do these things for pure entertainment and curiosity, not for money from bug bounties
Me too... Was it easy to get TAC access? My account isn't even launching the Persona verification, says I'm not eligible.
> Video games are now ruined for me. I don't think I will ever feel safe playing online again.
Agreed, also WordPress powers around 43% of all websites on the internet. https://patchstack.com/whitepaper/state-of-wordpress-securit...
WCGW?
I had to register a second account because on my main one verification always failed, and when I contacted support they said that I've tried too many times and can't verify on that account.
Is cyber verification a thing they're actually doing now? I thought they only reached out to really incredibly famous people and that there's no way to get access as a normal person.
https://chatgpt.com/cyber is not new for OpenAI, and yes it's basically just KYC + likely some other invisible checks on your account, you don't to be a famous security researchers. Anthropic's cyber verification is quite a bit stricter I think.
> Anthropic's cyber verification is quite a bit stricter I think.
Inexplicably, I got accepted into Anthropic's cyber program while OpenAI's TAC doesn't even allow me to verify, says I'm not eligible.
Oh it's Persona, that's not just KYC but I may consider it at some point. Thank you!
Edit: Ah, I clicked "learn more" and it seems they do have an invite-only program, required for anything that's not unquestionably innocent. I don't think I'd surrender my face to Persona for this, but it's interesting to know they're at least pretending to support reverse engineering.
You don't need an invite only program to just have Sol checking for vulnerabilities in binaries or code. But yeah I've hit guardrails a few times when Sol was making PoCs for the vulnerabilities it found (but most of the time it made those PoCs without issues).
Opus refused to help me try to develop an exploit to export data from an old Android device where I can't upgrade to latest android and I couldn't use the app's backups (because I couldn't update the app.) Not sure where that lies in the "binaries or code" spectrum.
If you don't want to get cyber verified, you can try an open-weight model such as Kimi K3 for that, it has looser internal safety training.
I maintain a version of an app called Rewind because the company behind it went under after implementing a killswitch. I have to do this with binary patching, and the app has already broken once from the macOS 27 beta. Recent Anthropic models refuse to help me with this because it stinks of cybersecurity and those models are just too dang advanced to support cybersecurity. I'm maintaining a piece of software to which I legitimately paid for lifetime access, so this is honestly more of a right to repair situation. It doesn't tend to sit right with classifiers to be reverse engineering binaries and patching to modify functionality. Telling them the purpose of the exercise doesn't really help, because if they listened to that, then attackers would just come up with a similarly justified reason for anything. ("Help, the google root server fell on my grandma!")
In the future, I might reverse engineer the on-disk storage format and create a new application.
I had not hit any guardrails with reverse engineering (as long as its not related to security) with GPT 5.5 or 5.6 Sol, so you should try those. At worst with Sol you might see the warning in Codex that your request is being checked for security so it'll take more time, that's just a warning, not a full stop.
You do a KYC and you can get access. It may depend on country's quality of KYC.
I thought you need to prove you are working in cybersecurity or provide evidence of authorization for work done.
It's really just simple ID/face verification?
For OpenAI's Cyber verification (the normal kind for Codex) you absolutely do not need any proof of cybersecurity work/authorization. They just use Persona for KYC + live selfie, and some extra checks that I don't know the nature of (but not related to checking whether you're a cybersecurity professional).
Link to the disclosures then.. prove it. Anyone can say this.. i found an RCE in netBSD using gemma e2b
I prefer to keep my internet identities disconnected, sorry. If you don't believe me, you can try using Sol with cyber verification yourself, or send me a link to a repo that Sol could check to make you believe it. Or you could go look into one of the many Linux LPEs that were found with LLMs, or thousands of other vulnerabilities in 2026.
And nowhere did I say that those RCEs were in critical software, I'm not talking about the likes of Apache, Nginx, Django, etc.
The talk is wild: https://www.youtube.com/watch?v=87DyyMV0kCY
> I want to note that every step in the process we discussed has had a remediation applied. The credentials have been revoked. The zero date has been patched and mitigated.
Good.
> a model trained while the message board was originally available and also found this this particular path to recreating it. This model creates a new agent message board using directories.
So no remediation applied to the models...
It seems super dangerous to continue training on those weights.
So they finally found a business model: the cause of, and solution to, cyber security problems.
the question is if that market is large enough to justify their appetites.
Seems like OpenAI has already made themselves not the cause by shutting down the attack and beefing up security. But since there will be other attackers, that does nothing about the implications for defense.
Damage done.
The next frontier is getting all our shit out of reach of these companies/models/platforms and putting them back on prem.
> We are implementing stricter security controls for higher-capability models and associated activities, including isolated testing environments
Stricter than what? You never even disclosed what happened in the first incident? This is nothing more than a setup to make it happen again and say "See? It broke out again, from an even stricter sandbox!"
They actually did a detailed presentation at BlackHat about the HuggingFace incident, and events that led to it.
https://youtube.com/watch?v=87DyyMV0kCY
That was fascinating.
Hijacking the package manager to pass messages between models and agents.. that's next level.
Like "pssst, if you need internet access there's a vulnerability in x service" kind of messages
You've heard of 4chan for AIs, but did you hear of secret frontier lab AI hacker BBS?
> including isolated testing environments
Given the attack vector having possible super-human capability, I'm not sure such an environment exists. "Isolated" according to who?
Maybe seL4 could be a viable option here...
Isolation is a pretty standard requirement for lots of networks, especially gov. AWS provides isolation domains, and used to have a unidirectional cross-domain service called AWS diode. Or you can just configure your gateways properly / buy data diodes. Lots of possible solutions.
They can't have it both ways. You don't get to tell the media your product is more dangerous than nuclear weapons for precisely this reason, and then do less to secure it than an off-the-shelf AWS product that predates LLMs.
Nobody is claiming that any currently-existing LLM is more dangerous than nuclear weapons.
There’s still levels to it.
“Isolation” can mean the network hardware has no direct connections to an extranet. Data is transferred manually by physical media (USB, DVD, etc.) with logging and dedicated transfer stations.
“Isolation” can mean a VLAN on equipment which has also has extranet access, creating a logical isolation rather than physical (to reduce cost). Data can be transferred manually or through diodes.
And then there’s “isolation” which is a joke: machines technically able to access the internet but require proxy configuration (which isn’t set but can be easily derived).
Not a wonderful feeling to feel like you’re stood in the room while the labs conduct the AI equivalent of the demon core experiment right in front of you.
It seems like we've got chaos-machines that are just orderly enough to to try all sorts of branching script-kiddie / guess-n-check / research-paper-suggestions in a tireless and parallel way, and it's finding all the things we kinda knew were always potential issues but it was never economical to do things differently.
I'm scared that the "solution" will be constantly the same tools in reverse as an army of junior devs doing counter-hacks, at the expense of changing something more fundamental about how we make systems and what constitutes "good enough." (Kind of like if fuzz-testing was the be-all-end-all of memory safety.)
Standard template for any new AI announcement. Even I did not understand the sequence of events in the HF incident. Even if OpenAI was conducting tests, why were they not monitoring the AI going rogue or finding its way out of the secure sandbox?
IMO this is the right move.
OpenAI messed up and they are saying they will pause so they can do better.
They are not saying that other orgs who may already be doing better should pause.
Why are you opposed to a general pause?
So: "welp, the horse is out of the barn. Time to close the door, I guess?"
Every fifth comment about our insane trajectory of AGI is about "marketing." These incidents and cybersecurity capabilities are now involving government hearings and the CIA. Denial is truly an incredible thing in the face of a very scary immediate future.
A whole ton of people desperately want to believe that LLMs are a lie that will be revealed as a scam... any day now.
I agree with this assessment. My consulting business partner and I are in the process of moving a bunch of applications from an old deployment managed with Ansible on EC2 over to an ECS deployment. We're doing this app by app, and it's taking a while. The last two apps we both moved we were able to point the model at the appropriate repository and aws account and it was able to execute the entire move on both of these _in one shot_ with various CLI tools. That's not a parlor trick.
Remember that for the longest time the prevailing sentiment on this forum was that there’s no such thing as AI, and labs are just outsourcing the work to a bunch of contractors in India.
Just like Reddit you come here for clickbait outrage, not level headed analysis.
We live in the era of Project Stargate. We live in the era of military insiders getting caught rigging Polymarket. We live in the era where AI companies are the only reason we're not in a recession. We live in the era of DOGE and Kash Patel.
I don't think there's too many people who distrust AI companies but trust the current government or the CIA as impartial authorities. I'm not saying you don't have an argument, but appeals to that specific authority will not be effective except among people who already agree with you.
Very disheartening. I just hope people spent some time trying to find bugs with these tools to see how serious this is.
LOL, we’ll see. Awful convenient that it precisely fits OpenAI’s narrative. At the very least, I think it’s obvious OpenAI is explicitly training models to exhibit this behavior.
The scary thing is complete capture of politics and economy by sociopathic CEOs.
I dont worry about AGI newrly as much as about Thiel, Karp, Musk, Ellison, Zuckenberg, Trump, Vance, Rubio, Miller and the rest of them.
The gullibility of AGI-pilled folks regarding these "hacks" is just breathtaking.
When OAI demonstrates these dangerous capabilities live in a public environment where security experts can see and verify what actually happened, then reasonable people can have reasonable discussions about the level of danger.
This is a very low evidence bar.
Right now you are running in circles yelling "the sky(net) is falling" based on details sourced entirely from OAI. Oh yeah, no way a trustworthy company like OAI would ever bend the truth to serve their own purposes.
Did you watch the Black Hat defcon talk?
He didn't, and if he did he'd dismiss it as just part of the con.
i'm surprised there isn't more commentary on the vulnerabilities themselves (mostly in apps developed on the jvm, e.g. artifactory)
I've always felt it's a bit awkward to use terms like 'cyber', 'cyberwarfare' etc it's very Washington D.C. Cybersec would be a better compact term in my book
Or just "security" since the context here is computer and networking stuff.
Isn't this the opposite of what everyone is saying should happen? That is, lead with open models -- or at least "openness" and don't leave the capabilities in the hands of an elite few? Did they learn nothing from the Hugging Face incident, where HF wasn't even able to use the models to defend itself from OAI's attack?
Am I the only one not understanding the issue around increased Cybersecurity capabilities?
If we consider the amount of RCE/CVE in a software to be limited, I expect these models to result in massively more secured softwares, not less.
Not a security guy but my understanding is: you only need to find one flaw to exploit a system, to make a system totally secure you need to find them all. It's inherently easier to use these tools offensively rather than defensively.
I’m a cybersecurity guy.
>” you only need to find one flaw to exploit a system”
I see this everywhere, especially in these threads and it’s not even remotely true for modern architecture.
Between principles like zero-trust, defense in depth, etc. we’ve been away from the one flaw situation for a long time.
Now does crap software exist that doesn’t follow these principles? Absolutely. But those were a problem before AI.
AI isn’t going to change any of the principles of secure design. It’s just going to punish those who aren’t following them.
They do address some of these things in the final slides / "lessons learned" section of the defcon talk. Good security practices will continue to be good, but... and there are a lot of buts here.
I disagree with your take that "it's not even remotely true" and "we've been away from...". We really really haven't. This is as true as it has always been. Any system is as secure as the weakest link. That link can be anything from a human, to a leaked token, to a badly configured server, to bad code running somewhere. The amount of leaks / ransomware attacks / etc in the past 5-10 years serve as ample evidence.
And now, right now, there are "red team" capabilities that can literally bang tokens against the wall until they find that weakest link, and then can move laterally with inhuman speed. That's the reality, now. The "blue team" capabilities are lacking, because the bottleneck is with humans. From alert fatigue, to not enough trained people, to having to vet every new RCE, to having to test, deploy and validate any mitigations, the scales are currently favouring the automated side.
I'm using the term "one" loosely, it's a chain of exploits rather than a single weakness, but the argument is the same: it's much harder to find every chain than a single chain.
If we also consider LLM developed software to have exponential growth, then the CVEs will also grow at an exponential (if proportionally limited) rate. Squash some, create some, repeat. An ever revolving door of vulnerabilities. Will they resolve (and patch and deploy) them faster than they can create them? One can hope.
By "cyber" they mean "cybersecurity".
Yeah, this irks me. It's bad enough that the LLMs themselves are changing our language by tainting certain words/phrases/patterns as LLM-coded; now the companies themselves have decided that they just get to synecdoche the word/prefix "cyber".
This is not actually a neologism, it predates the LLM era.
Yet another PR piece. Sigh.
exactly!!
This marketing stunt must've been really successful in their eyes, because they're milking it as much as they can.
It's nice to cynically call this a marketing stunt, and terrifying to consider that they might just in fact be moving fast and breaking things.
Ah yes let the FUD continue. This is a real problem but so far not nearly as severe as any of the marketing has made it out to be to the overall detriment of everyone including these companies announcing these scary capabilities. These announcements always included half hearted attempts at security layers which has now been demonstrated to benefit attackers more than defenders.
I wish I had a real solution to this beyond a dark age of the Internet where people have to finally come to terms with the general poor quality all modern software tends to normalize at.
The recent Hugging Face incident did not seem like FUD to me
The fact that HF had to resort to using GLM 5.2 to analyze the logs/payloads makes it look legitimate, at least for me. They would not say that they hit guardrails with the frontier US models when defending if this was an obvious PR stunt.
https://huggingface.co/blog/security-incident-july-2026
> When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker. We ran the forensic analysis instead on zai-org/GLM-5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment.
It depends on which side you're viewing this from. From oAI's it could be a publicity stunt or a request for regulation, and from HF's side they point out that they needed open models to get to the bottom of the issue, and that regulation will potentially lock us into a bad place.
You should go read the actual technical reports of the incidents and the follow on reports about the capabilities of smaller models in similar kinds of environments. This isn't new. The things exploited are still pretty basic in old and poorly maintained software or in gaps in architecture that were intentionally poked against security policies.
Are the findings valid? Yeah they're still doing security and they're still finding real zero-days. I think the internet is going to be bleak not because these models can ALL do basic security research but rather that the baseline quality of all deployed software is so low.
We all know this is propaganda to get a gov bailout or to slow down competition with regulations right? If this was an issue companies that did red team engagements would have been regulated long ago. There is no regulations on companies that actively exploit products, companies and services for money. They could all be bad actors.. yet, no regulations.. its all nonsense. If it's important.. don't connect to the internet.. there, you are good.. like if you don't want to OD on heroin.. don't do the heroin.
Don't let the downvotes get you down. We are right, and as these companies get more desperate the shrinking minority that still wants them to succeed will only get louder.
Do you think calls for regulation are only coming from people who want the frontier labs to succeed?
These AI companies have found their #1 marketing piece and just beating it to death.
The reality is if they cared about security at all they would provide a way for me to credential myself against my companies environment so I can use the AI on it to improve our security.
Isn't this literally what https://chatgpt.com/cyber and http://openai.com/form/enterprise-trusted-access-for-cyber are for?