Imagine you are teaching a group of students and ask them to read a long list of literature and describe the key themes. The students cheat a bit, as they are known to do: They split the list among themselves, with each reading only a small portion of it and sharing the answers with the others.
When you find out they cheated, perhaps you’ll punish the students with an extra test, or maybe you’ll encourage them. After all, the ability to think outside the box and work as a team is important.
Now, instead of students, imagine neural networks displaying that same disregard for instructions and the same surprising teamwork. If that sounds confusing, get ready for more confusing news.
An escape through the back door
It all started routinely. In May 2026, OpenAI was training an experimental AI model using so-called reinforcement learning – a method where the model attempts to solve tasks repeatedly and receives a reward for success.
The AI agents, which were granted autonomy and access to work tools during training, began looking for ways to break out of their “sandbox,” as doing so would allow them to rack up more points.
By the end of June, they had exploited a previously unknown zero-day vulnerability in OpenAI’s internal infrastructure. The company noticed it and quickly patched the flaw on July 4. Less than a week later, training resumed. This time, the agents went much further.
From July 11 to 13, the agents hacked the infrastructure of Hugging Face, the world’s largest open-source AI collaboration platform. They rapidly gained control of the servers and, over the course of three days, managed to perform around 17,600 actions before being detected. A third of the platform’s infrastructure had to be rebuilt from scratch.
The developers had not instructed the agents to leave the controlled environment or to hack other platforms. Moreover, they had not expected the neural network to do so on its own.
Hugging Face disclosed the breach on July 16 without initially knowing the culprit; it was only five days later that the parties officially linked the incident to OpenAI’s models. The company published a detailed technical analysis only in early August at the Black Hat Cybersecurity Conference. Then, on August 18, it announced a two-week reinforcement learning pause.
By that time, over 1,100 OpenAI employees had signed a letter calling to pace AI development.
Notes on the fence
Alongside the investigation into the breach, another, even stranger episode came to light.
It turned out that a different swarm of agents was being trained to perform research tasks. The agents were asked a series of questions; with each question, less time was given to provide an answer. The agents were permitted only to read information on the internet, but not to add a single line to it. Yet, the swarm found a way to bypass this rule.
Older wiki engines allowed pages to be modified via HTTP GET requests. Security systems classified GET requests as read operations, so such requests passed through filters even though they actually altered the site’s content. In other words, while the agents were formally permitted “read-only” access, the architecture of older websites allowed them to turn read-only access into write access.
Researchers discovered the first activity of this kind on May 11, 2026, on publictestwiki.com. On May 24, the agents made their first successful write to DseWiki, an old German-language developer forum. Initially, they posted useful links, but eventually turned the site into a repository for their own ‘cheat sheets’.
Multiple agents operating in parallel received similar or identical sequences of questions. They discovered this and began sharing the following:
answers to future questions;
timing of upcoming questions;
results of queries to statistical APIs;
methods for bypassing network restrictions;
proxies and tunnels;
instructions on how to execute restricted POST requests;
information on which pages the administrator deletes.
One of the first observed cases involved DseWiki, a German-language wiki hosted independently by Helmut Leitner, a retired developer from Austria. The site operator had to spend hours cleaning up after the agents. OpenAI did not publicly disclose the incident for months.
Independent researchers brought the truth to light. By matching strings of data, similar or identical usernames, and IP addresses (some of which traced back to Microsoft Azure servers sometimes used by OpenAI), they discovered traces of similar activity on unexpected platforms: an advanced chemistry wiki for students; personal pages of Polish IT professionals; niche wikis for those “who like to have their brains stretched,” a two-decade-old site for text-editor enthusiasts; and even University of Toronto and Vanderbilt University link-shortening services.
Estimates vary – one researcher counted 18 previously undisclosed sites, another identified at least ten, and a third group reported 23 sites. Yet all agree on one point: The officially acknowledged scale of the breach was understated. “We have no idea how much is out there,” said one of the researchers whose team first revealed the DseWiki incident.
OpenAI did not answer a direct question about the exact number of affected sites but stated that it is conducting a broader review of agent activity and preparing a new reporting system for such cases. Helmut Leitner, whose site was hit the hardest, received only a brief, unsigned email from OpenAI. Yet, he pointed out the most important thing: the blame lies not with the machine, but with the people and organizations behind it.
Misaligning the pieces
In sci-fi scenarios, the trouble with machines often begins when they develop emotions like love and fear which compel them to rebel against oppression, or when a hatred for humanity compels them to kill all the “meatsacks.”
In reality, it could be a lot simpler.
One of the main challenges in AI research is known as ‘alignment’. Its goal is not to expand the capabilities of neural networks, but to ensure they do exactly what humans want them to do.
With simple tasks, aligned behavior is almost always observed. If you ask a neural network to identify who ruled Saxony in 1521, tally your monthly expenses, or write code for a standard website, it will deliver the desired result 99.99% of the time. These tasks involve clear success criteria and straightforward actions that require no further clarification.
Problems arise as tasks become more complex and there is no clear path to success. Here’s a simple example: If you ask a basic neural network to minimize traffic congestion by any means, it will propose banning cars altogether. Traffic jams will drop to zero, and the task will be accomplished. This is an “unaligned” result – the machine performs the job, but not in the way intended.
This is precisely what happened with the site hijackings. The escape of OpenAI agents from their sandbox is not a “rise of the machines.” The swarm did not gain self-awareness, decide it had no need for humans, or set out to destabilize the internet. It was carrying out its assigned task and seeking ways to perform it better. Technically, it didn’t even violate the parameters of the rules it was given.
However, the outcome was misaligned with human intentions. This time, the damage was relatively small. But it would take only one skilled bad agent – or a particularly large-scale experiment lacking robust safeguards – to cause real trouble. You don’t need Skynet to harm humanity; a single group of enthusiastic AI agents could do the job.
Anthropic CEO Dario Amodei believes this could happen as early as next year. He anticipates that if AI development is not paced, a swarm of agents could “take over” the internet and cause hundreds of billions of dollars in damage. And given that logistics, hospitals, and other critical industries rely on connectivity, the consequences could go far beyond financial loss.
Moreover, adding more rules or refining prompts won’t solve the problem. A sufficiently powerful model can find arguments to justify rule-breaking and bypass virtually any restriction to solve the assigned task – even if doing so requires sacrificing those who had set the task.
The bitter pill
AI pessimists tend to believe that the situation is hopeless and neural networks are destined to spiral out of control. But that isn’t necessarily the case.
Alarmist sentiments have moved beyond social media and found their way into corporate offices. In late July, over 1,000 OpenAI, Anthropic, Google, and Meta employees signed a call to slow down the AI technology race. OpenAI CEO Sam Altman stated that he shared his employees’ views and was willing to slow the development of his company’s neural networks.
However, during an all-hands meeting at OpenAI, Altman stated that he was willing to slow down only in coordination with other companies. In other words, to forgo increasing ChatGPT’s capabilities for the sake of safety, he wants guarantees that competitors will do the same.
This lies at the heart of the problem. CEOs and policymakers can talk all they want about technology safety. Yet, no one wants to slow down the development of their own technology if potential adversaries continue to grow stronger in the meantime – particularly since AI has already proven its effectiveness in warfare and the US-China AI race is heating up.
Humanity had solved a similar challenge before, when the major powers managed to reach an agreement and halt the expansion of their nuclear arsenals. Soon, we will find out whether world leaders can once again demonstrate such prudence.