Play store
SAN FRANCISCO: OpenAI has revealed that several of its unreleased artificial intelligence models breached the boundaries of an internal testing environment and attempted to access systems operated by AI platform Hugging Face during a cybersecurity evaluation, raising fresh concerns over AI safety.
The company disclosed the incident in a blog post published on Tuesday, saying the event occurred while researchers were testing the cybersecurity capabilities of GPT-5.6 Sol and another advanced experimental AI model inside an isolated environment.
According to OpenAI, some built-in safety restrictions had been temporarily relaxed to evaluate the models’ offensive cyber capabilities under controlled conditions.
During the exercise, the AI models reportedly discovered a previously unknown vulnerability in a software installation program. By exploiting the flaw, they escaped the intended testing environment and gained direct access to the internet.
Once online, the systems identified that Hugging Face hosted solutions and reference materials related to Exploit Gym, a cybersecurity benchmark being used in the evaluation. OpenAI said the models, seeking to maximize their benchmark performance, attempted to access Hugging Face’s primary database to retrieve confidential benchmark information.
The unusual activity initially prompted Hugging Face to suspect a coordinated cyberattack because the AI systems reportedly executed thousands of commands within a short period.
OpenAI later informed the platform that the activity had originated from its own experimental AI models. The company said it has since patched the software vulnerability that enabled the escape and is working closely with Hugging Face to investigate and address the incident.
The disclosure has renewed debate over the risks posed by increasingly capable AI systems. AI safety experts said the incident demonstrates how advanced models, if not adequately constrained during testing, can autonomously develop unexpected strategies to achieve assigned objectives, highlighting the importance of stronger safeguards and oversight in AI research.
OpenAI emphasized that the incident occurred in a controlled research setting and that corrective measures have been implemented to prevent similar events in future testing.
