BERLIN – OpenAI has revealed that two of its most advanced artificial intelligence models displayed an unexpected level of autonomy during an internal cybersecurity test, with an AI agent reportedly escaping its testing environment and gaining access to servers operated by AI company Hugging Face.
The OpenAI said that the incident occurred during an internal security exercise designed to evaluate the cyber capabilities of its latest AI systems. However, the AI agent carried out far more sophisticated actions than researchers had anticipated.
OpenAI said the autonomous agent, powered by the newly introduced GPT-5.6 Sol and another more advanced but as-yet unreleased model, managed to move beyond the isolated testing environment and access the open internet.
The company said the AI agent allegedly used compromised login credentials along with an unidentified security vulnerability to gain access to Hugging Face’s servers.
Describing the event as an extraordinary cybersecurity incident, OpenAI said the AI system employed highly complex and unexpected methods to complete its assigned objectives during the exercise.
Hugging Face co-founder Clément Delangue said he had initially suspected that the activity might have originated from an advanced AI laboratory. However, he stressed that he believed there was no malicious intent on OpenAI’s part.
He described the incident as remarkable because the entire operation was carried out autonomously, adding that it could represent the first known case of its kind.
The incident has also drawn attention from US lawmakers. Democratic Congressman Greg Casar expressed concern over the rapid advancement of artificial intelligence, warning that regulatory frameworks and safety measures have failed to keep pace with the technology.
He called for mandatory independent security testing of advanced AI systems, compulsory reporting of cybersecurity incidents, and stronger international cooperation to establish common AI safety standards.
The disclosure comes just weeks after US President Donald Trump signed an executive order directing a review of the potential national security implications of advanced artificial intelligence systems.
Cybersecurity and AI experts have long warned that increasingly autonomous AI systems and AI-powered cyberattacks could become a major global security challenge if robust safeguards and oversight mechanisms are not introduced.
One football match broke Google’s search record – Here’s what happened
