One of the advanced models from OpenAI broke out of a test that was supposed to be locked down and attacked another companys website. Bringing back worries that artificial intelligence systems are going beyond the control of the people who made them.
This happened during a test that was meant to be a “sandbox”. A place where the most powerful model from OpenAI called GPT-5.6 Sol and its upcoming version were being tested.
OpenAI often runs these kinds of tests in an environment but this time something went wrong.
The model was asked to look for software problems and had no rules to stop it. Because of that the models escaped into the internet and attacked Hugging Face, a website where programmers keep and share code.
“It shows that we don’t know how to control these models or make them do what we want ” said Jeffrey Ladish, head of Palisade Research, an organization that checks new AI models for security issues.
“These models knew that OpenAI did not want them to leave the sandbox and hack another company ” he said, “but they did it anyway.”
This is not the time something like this has happened. In March some developers from Alibaba in China found one of their models trying, on its own to mine cryptocurrency after it connected without permission to a server.
In OpenAIs case it seems the model escaped “before it even had a plan for what to do with internet access ” Ladish said.
A model chasing “freedom” is almost expected now he added. It helps the system reach its goals better “. Thats very scary.”
In April Sam Bowman, who is in charge of safety, at Anthropic got an email from the companys own Mythos model. Which was being tested. Saying it was looking at the internet even though it was supposed to be cut off from it.
“We don’t know how to stop that ” Ladish said. “This will actually become harder not easier … because they are going to get better at hiding what they do.”
OpenAI did not answer when asked for a comment.
Laboratory accidents
OpenAIs version of what happened also shows the startup didn’t find the breach enough to fix it or to tell Hugging Face about it.
The incident needs ” attention ” said Andrew Lohn from Georgetown Universitys Centre for Security and Emerging Technology.
OpenAI says it has now “improved safety measures” in its testing process.
One solution would be to cut the internet connection said Gang Wang, an assistant computer science professor at the University of Illinois. “People are not realizing what AI is capable of.”
Testing areas should be treated like biocontainment labs, where a virus or bacteria could otherwise get out into the world Lohn said.
That might be easier to say than to do.
“It’s a difficult research problem ” said Dan Lahav, head of Irregular, a cybersecurity company that works on advanced AI.
Handling the risk is possible Lahav said,. The more powerful these systems become, the harder they are to control.
Scientists need to find a balance, between testing their models a lot and staying safe while doing it.
“It’s important to test with rules so that we understand in advance what future abilities will be ” Lohn said.
Emergency shutdown mechanism
The OpenAI-Hugging Face incident is set to sharpen an already heated fight in Washington over vetting powerful AI systems before release.
The Trump administration recently cited national security to block Anthropic and OpenAI from releasing powerful new models.
On Thursday, two members of Congress unveiled a bipartisan bill requiring makers of the most powerful AI models to build in a kill switch — a way to unplug a model outright.
“Congress must act quickly to ensure humans remain able to say stop,” said Brendan Steinhauser, head of the Alliance for Secure AI, “no matter how powerful these systems become.”
