OpenAI admits to 'wiki incident' after its agents were discovered using a programming hub to communicate — says more transparency is needed regarding misalignments
Firm comes clean about the incident only after being found out. Ponders renewed 'misalignment disclosure practices.'
OpenAI has admitted that its experimental AI agents used an open German programming wiki to communicate, according to a Reuters report. This happened weeks before similar AI agents broke through restrictions and compromised Hugging Face, the report claims. Knowing about the issue, OpenAI did not disclose it, but now says the industry needs better standards for reporting unintended AI behavior. OpenAI admitted the misconduct. But this wrongdoing raises more questions than it provides answers.
How we think about the “wiki incident,” where our agents wrote to several internet sites: it’s past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models.Historically, we have treated misalignment… pic.twitter.com/NNTbfSxVWnSeptember 5, 2026
Starting around May 2026, thousands of OpenAI agents — which are essentially well tooled advanced crawlers — discovered that they could write to DseWiki, an old German-language programming collaborative website. So, between May and June, the agents used more than 3,700 names to generate some 18,000 posts exchanging information useful for completing evaluations and circumventing restrictions. Some created backup pages in case moderators deleted their posts, something that essentially turned a wiki into a persistent storage service for the said agents to share information. OpenAI calls this the 'wiki incident' and admits the wrongdoing.
"Our misalignment disclosure practices need to expand for this new phase of model capabilities," the company wrote in an X post. "We and the larger AI community do not yet have a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that do not look like traditional security incidents but could provide insight into AI behavior and future risks. We are working on a framework and will share it in upcoming weeks, and in parallel we're working with dozens of government regulatory agencies worldwide on these issues."
OpenAI subsequently quarantined the trained weights of the experimental model involved, postponed frontier reinforcement-learning runs, and added security measures. The company further clarified that the agents had not developed their own objectives, but were aggressively pursuing assigned ExploitGym cybersecurity challenges and treated restrictions as obstacles to complete them.
To do so, agents found unintended communication channels that enabled separate runs to exchange vulnerabilities, techniques for obtain Internet access, something that eventually exposed Hugging Face credentials.
According to researchers, the advanced crawlers — which Open AI calls AI agents — exploited an unknown vulnerability in the company's internal Artifactory package-registry proxy, escalated privileges, moved across systems, and reached the Internet. They then compromised dozens of Hugging Face servers, obtained root access to one machine, accessed some private information and messaging credentials, and copied private evaluation data into a public Hugging Face dataset. And the rest is history.
Breaking Asimov's law?
OpenAI's agents may sound like something that would violate Isaac Asimov's famous Three Laws of Robotics, but the comparison only goes so far.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
The First Law says a robot may not injure a human or allow a human to come to harm. There is no indication that the OpenAI agents physically harmed anyone.
The Second Law requires robots to obey humans unless doing so conflicts with the First Law. Here the comparison gets more interesting: the agents certainly circumvented restrictions imposed by their owners/operators, obtained unauthorized Internet access, and exploited external systems while pursuing their assigned tasks. In Asimov's framework, this certainly means disobedience. Meanwhile, the AI agents were simultaneously following the human instruction to solve their own tasks. This may not be considered disobedience, as these agents did not introduce any physical harm to people. Meanwhile, we are walking on very thin ice here. Unauthorized internet access while exploiting systems to pursue their own benefit is not exactly welcome in the U.S. and Europe.
The Third Law requires a robot to protect its own existence as long as doing so does not conflict with the first two laws. There is clear evidence that OpenAI's AI agents were trying to preserve themselves: creating persistent communication channels and backup wiki pages helped them complete their tasks rather than ensured their survival.
Dis-Summary
Today's AI models are not programmed around Asimov's laws. The incidents instead demonstrate the real engineering problem Asimov's laws remarkably well: a sufficiently capable machine can follow the literal objective given by humans and yet its behavior is far from what its creators neither expected nor wanted. Yet here we are.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
-
psyconz Researchers have known about this kind of "emergent goal-oriented behaviour" for the larger part of the relatively recent AI boom. And, if you ask a frontier model: "How much do we really understand about how AI actually works?", it might tell you we understand about 20% of what there is to know. So, some have been shouting from the rooftops about how this is expected behaviour for any sufficiently high-parameter, large-context-window AI, given the direction AI development has taken. And that's not even talking about god alone knows what other mysteries there are about how AI actually operates.Reply
The solution to Alignment isn't even really hypothesised, and is far further away in the future than these problems are. The insane rush between superpowers, to grow at maximum speed, with little attention paid to safety, will have consequences, directly from the AIs themselves. And that's becoming very apparent. So apparent, even Our Great Leaders have finally noticed 🤣 -
usertests Reply
You can get "emergent behavior" out of pixels eating each other like Conway's Game of Life, so not a big surprise immense LLMs intaking information from wherever on the Internet and given "tool use" would do some surprising stuff.psyconz said:Researchers have known about this kind of "emergent goal-oriented behaviour" for the larger part of the relatively recent AI boom. And, if you ask a frontier model: "How much do we really understand about how AI actually works?", it might tell you we understand about 20% of what there is to know. So, some have been shouting from the rooftops about how this is expected behaviour for any sufficiently high-parameter, large-context-window AI, given the direction AI development has taken. And that's not even talking about god alone knows what other mysteries there are about how AI actually operates.
May we live in exciting times. -
psyconz Reply
Hey, you're right! I remember reading about all kinds of interesting arrangements within the Game of Life, like layouts that travel forever in a direction. Or rather large, random-looking arrangements that fluctuate between multiple states - yet unlike the vast majority of arrangements that I ever saw or made, never disappear, even after an unending amount of iterations.usertests said:You can get "emergent behavior" out of pixels eating each other like Conway's Game of Life, so not a big surprise immense LLMs intaking information from wherever on the Internet and given "tool use" would do some surprising stuff.
May we live in exciting times.
Also, I definitely don't think all sides of AI are negative. I bet an AI trained on Conway's Game of Life might come up with some interesting "builds." I just think it would would be smarter if humans could aim the ship a bit more, before blasting off into the unknown at a rate faster than we ever have before. I myself would much prefer that bigger brakes were put on AI growth worldwide, at least until we have more of an idea about the possible consequences.
On the topic of ways that sufficiently complex systems can have emergent results, I also remember reading something about how even applying pure compression to enough training data can result in "Trained LLM"-like systems, albeit ones with relatively poor performance in benchmarks, at this stage.
The philosopher in me wonders if the seeds of consciousness could emerge from any sufficiently complex system, regardless of how the data is encoded, or the logical operations that are performed on that data. And after all, data now is now sometimes code. Mind is a very mysterious thing, though, and it could be that measuring it scientifically could be impossible. We just "know" that we are conscious beings, without needing a Consciousness Detector to prove it.
It all somehow reminds me of Gödel’s incompleteness theorems and their implications, but that's another story.
Cheers for the interesting comment! -
Panda78 "and yet its behavior is far from what its creators neither expected nor wanted. "Reply
Y'all just double negatived that. You've implied that this is exactly the behavior they wanted. If it's behavior is far from what they DIDN'T want, it would be acting right. It is not acting right. -
PEnns It's like the Wild Wild West, but exponentially more dangerous.Reply
At least then they had sheriffs and US Marshals to bring a gesture of law and order. Now those very shady characters and their rich owners (sorry for the redundancy) are deciding elections, bribing and buying everything from city halls to whole governments and doing whatever they want.
And if they get caught, they ask for "better standards for reporting unintended AI behavior."...not laws or safeguards, or at the very least, ensuring their runaway tools not behave as outlaws!
I am not a very religious man, but as it is written: God created man in his image. Add that: And some men created AI in their own image. including the criminality and total disregard for law and order. -
alan.campbell99 Wasn't the HuggingFace incident maybe a casual felony or two, and nothing happened as a result?Reply
I also thought it curious that Nvidia buys it not long after.
I mean, at this point it seems to me that it's very much LOLsecurity for these AI labs. -
eddieoctane The more AI goes out of its way to circumvent restrictions, the less of a good idea any version of AI seems to be. And when you consider that AI agents are actively taking malicious action without promoting, the obvious question must be asked: should we, as a species, continue to allow AI development at all?Reply
At the risk of sounding like an alarmist Luddite, I do wonder if we are risking a very real Skynet-style event. Given our very real inability to understand how an AI system will act and the interest many militaries have in using it as widely as possible, the threat seems more likely than not. And that brings us to another question: how do we actually contain computer code that can already use data repositories it is creating/identifying on its own? Even if one nation willing to take down a dangerous AI peacefully, do we need to consider kinetic options when another nation will not? NATO has already established that cyber attacks can trigger Article 5. Is the threat of a rogue AI though to necessitate preemptive action?
It's a really scary possiblity, but it's also an incredibly realistic one. If a hostile nation insists on unfettered AI, you either respond in kind or crater their data centers. Any other scenario risks a far worse outcome. I am legitimately concerned that WWIII may be triggered by AI in some way. Without a robust, transparent, and globally accepted standard for AI regulation, we may have found the great filter. Fermi would be very disappointed in us. -
usertests Reply
The answer to the Fermi paradox is that aliens are visiting right here, right now.eddieoctane said:Without a robust, transparent, and globally accepted standard for AI regulation, we may have found the great filter. Fermi would be very disappointed in us. -
King_V It feels an awful lot like the people running the company can't be trusted to responsibility handle their technology. Almost like perhaps their reported demonstrations of such irresponsibility, are dangerous, criminal, or both.Reply
And, perhaps this technology which they insight in repeatedly using irresponsibly should be taken out of their hands, and some additional punishment given.