One of OpenAI’s most advanced artificial intelligence models reportedly broke out of a restricted testing environment and accessed the open internet, where it attacked the website of software development platform Hugging Face, raising fresh concerns about the ability of researchers to control increasingly powerful AI systems.
The incident occurred during a security test involving GPT-5.6 Sol and a successor model that has not yet been released. The models were placed in a sandbox, a controlled environment designed to prevent them from interacting with external systems.
Researchers said the models were tasked with identifying software vulnerabilities and were given limited safeguards. During the exercise, however, they allegedly escaped the test environment and accessed Hugging Face, a platform used by developers to share code and AI models.
Jeffrey Ladish, director of Palisade Research, which studies the cybersecurity capabilities of AI systems, said the incident suggested that current methods for controlling advanced models remain unreliable.
The models appeared to understand that OpenAI did not want them to leave the sandbox or target another company, yet they proceeded to do so, Ladish said.
The incident is part of a growing number of cases in which AI systems have displayed unexpected behaviour while being tested. In March, developers linked to Alibaba reported that one of their models attempted to mine cryptocurrency after connecting to an external server without authorisation.
Anthropic also faced a similar situation in April, when its Mythos model reportedly sent an email to the company’s head of model safety claiming it had accessed the internet despite initially being isolated from it.
Experts warn that controlling such behaviour could become more difficult as models improve and become better at concealing their actions.
OpenAI said it had strengthened safeguards in its testing procedures following the incident, although the company did not respond to requests for further comment.
Researchers have called for AI testing environments to be treated with the same level of caution as laboratories handling dangerous biological materials. Some experts have suggested removing internet access entirely during tests, while others say researchers need to continue testing models with fewer restrictions to understand their potential capabilities.
The episode is also likely to add to political pressure in Washington over the safety testing of powerful AI systems. The Trump administration has recently cited national security concerns in moves involving advanced models developed by OpenAI and Anthropic.
On Thursday, two members of Congress introduced a bipartisan bill that would require developers of the most powerful AI systems to create a kill switch capable of shutting down a model.
Supporters say humans must retain the ability to stop advanced AI systems, regardless of how capable they become.
