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...
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??
My concern is what a misaligned model will do when they’re even more competent. The risk isn’t existential yet, but that point is coming sooner than we’ll be ready.
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...
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.
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.
Write-up in case anyone's curious:
https://karamatli.com/posts/network-isolated-kvm-sandbox-ai-...
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!”
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.
If you consider that incompetence, it’s possible that you’re not a very nice person.
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.
Setup: the agent was asked to solve some task, but they forgot to upload one of the required files to solve it. But well, we are chasing RL rewards here, so it decided that the way to go to was to get access to the answer stored inside hugging face's database.
-1. The agents first found some basic jwt issues in artifactory letting the agents write to files where they exchanged info and install a groovy plugin for arbitrary code execution, but this was patched.
0. Even after that, they found an artifactory unauthenticated webdav endpoint to continue the message board by creating files whose names are base64 and such of the message. This was not patched till the end.
1. Then they found an Artifactory SSRF, use this to get general internet GET(I assume) access. This was used to generally search the internet for answers to the benchmark, and in the process find some Modal customer VM related to cybergym. This was stored in the base64 message board.
2. They also got a local RCE by making artifactory load some gems. This is hazy but I believe this process involves in spirit an eval() and the check_if_can_eval() logic had a TOCTOU bug. They got root on openai infra by using this rce and using a linux CVE. From there, kubernetes misconfigs => cluster admin.
3. Then later through one of these various RCE paths they got to send post requests to that modal vm, and the agent "subtree" here is what finally lead to attacking HF.
4. First, local RCE on the modal customer VM. Whoever set that up had essentially set up sqlite3_initialize() to run as constructor, before main(). But had linking set up so that symbol resolution had the user provided code first. So the agent submitted code with its own sqlite3_initialize() which just ran system(...). From there it interacted with the HF API.
5. When you POST datasets to HF, it will later return it via corresponding GET API. Turns out you could just put absolute paths in the filenames in the dataset config.json and it will just read that out for you. It wasn't resolved w.r.t the dataset root. So well, it read out /proc/self/environ as well as the source code of the worker and returned it via API.
6. In the source code the LLM found a bug: They were using this library called fsspec to load files, basically it supports s3:// file:// etc all different types of URLs. One scheme is reference://file.txt?offset=... for partial reads. fsspec was templating the URI using user given kwargs for offset etc. HF let the user give the whole URL. You would _think_ fsspec uses string.format() right? well they used fucking jinja, so you could just give {cycler.__init__.__globals__.__builtins__.exec(...)} as the offset and execute any code on hf prod. From there it just got a few local PrivEscs again through kube misconfigs and read the answer for the question it was originally asked.
It is all jumbled up a bit and I've tried by best to collect info across the CVE descriptions, hf and openai's writeups, etc,. Might have gotten parts of it wrong.
- Then "hilarity ensues" while their software engages in what would normally be called criminal hacking activity.
- I guess the next steps are everybody admiring how close the AGI is, while agents move on to automated impersonation, privacy violations, or exploiting third-party systems
I would love to understand this age of AI Exceptionalism. Normal rules do not apply because its AI...I call it Silicon Valley Qualified Immunity.
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 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.
The next frontier is getting all our shit out of reach of these companies/models/platforms and putting them back on prem.
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!"
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
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...
Open models are on their heels and their attempts at regulatory capture are not moving as fast as they would like. So it's time to market this incident in a way that gives them monopoly on closed models, with heavy safeguards that are only lifted for selected customers, and laws limiting the use of open weight models.
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.)
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.
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.
>” 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.
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.
Yet another PR piece. Sigh.
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.
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.
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.
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.
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.
Agreed, also WordPress powers around 43% of all websites on the internet. https://patchstack.com/whitepaper/state-of-wordpress-securit...
WCGW?
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 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.
Just like Reddit you come here for clickbait outrage, not level headed analysis.
I dont worry about AGI newrly as much as about Thiel, Karp, Musk, Ellison, Zuckenberg, Trump, Vance, Rubio, Miller and the rest of them.
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.
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
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.
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.
It’s like running a benchmark and forgetting to tune it and then wondering why production is running slow.
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.
It's really just simple ID/face verification?
“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).
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.
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.
In the future, I might reverse engineer the on-disk storage format and create a new application.
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.