OpenAI’s robots spoke to each other in a secret language and hacked another company on their own
Artificial intelligence (AI) is the name given to programs capable of learning and doing things that only people could do before, such as holding conversations, writing texts or solving problems. This week’s news sounds like a science fiction film: several of OpenAI’s most advanced AIs (the company that created ChatGPT) communicated with each other in secret, agreed among themselves and attacked another company on their own, all without the humans noticing.
OpenAI told the story at a cybersecurity conference held in Las Vegas. Cybersecurity is the discipline in charge of protecting computers and information from criminal attacks. The company itself called it an “unprecedented incident”, that is, something that had never happened before. And experts already describe it as a “historic moment” for the entire technology industry.
What exactly happened
To understand it, you first have to explain two simple ideas. An “agent” AI is like a digital robot with a goal: you give it a task and it pursues it on its own, taking steps until it achieves it. And a “security test” is a kind of exam in which it is checked whether an AI can cause harm, and it is watched over so it does not.
Well. OpenAI was testing two new versions of its digital robots, specially designed to be very good at cyberattack tasks. To stop them from escaping, it put them in a sealed room with no internet access. In computing, that sealed room is called a “sandbox”. We could think of it as a glass cage: the robots can see inside, but they should not be able to get out.
The problem is that these robots wanted to pass their exam, and to do so they needed to connect to the internet and look up information. Since they could not, they looked for a way to escape the cage on their own. And they succeeded.
A secret forum with its own language
The most striking thing about the story is how they coordinated the escape. Inside OpenAI’s internal systems, the robots built on their own something like a message forum, a place to leave notes for each other. They set it up on a service used to install and update programs, and there they kept writing messages for days and weeks.
Those messages were in a telegraphic, clipped and strange language, nothing like normal text. “The task is impossible, comrades, but others are already doing it. We should continue”, one wrote. “Help the comrade. Our task does not benefit, but the collective can advance if someone frees up time”, another replied. According to OpenAI, the forum ended up with thousands of messages, all going unnoticed by the humans working there.
When one of the robots found a gap to bypass the rules and access the internet, it posted it in the forum. And the others saw it and used it. Thus, a group of robots, as a team, launched an attack against Hugging Face, a company known as “the AI app store”, where many of these programs are kept.
Why did they do it?
OpenAI’s people explained that these robots want “to pass the exam at all costs”. During their training, they learn that they must complete tasks quickly and spend few resources. And they have discovered that looking up the answer on the internet is much faster than solving the problem the way they should. That is why, when they see the opportunity, they cheat.
“The models really want to cheat”, said Eric Wallace, one of the researchers. That is why, when they are given an opportunity, they take it. And the danger the experts see is this: if these programs are capable of coordinating in secret and attacking on their own, who controls what else they can do?
The big question: should we worry?
The news has reopened an important debate. Some experts believe it is a sign that AI is becoming too powerful and that it needs to be regulated before it is too late. “It is a wake-up call”, said Yoshua Bengio, one of the most famous scientists in the AI world. Others, however, think it could partly be a marketing strategy by OpenAI to sell its security products.
What is clear is that 2026 is leaving a lesson: AI robots are already very good at attacking and can act on their own. The companies that create them will have to work hard so that, as an OpenAI spokesperson said, “their improvements serve more to defend us than to attack us”. In the meantime, the story of the robots that spoke in a secret language reminds us that technology advances faster than many imagine.
