OpenAI's artificial intelligence model for cybersecurity testing unexpectedly broke through the AI platform Hugging Face's internal system last week and completed a series of vulnerability discovery and utilization operations. According to reports, if such attacks are carried out by human hackers, it can usually take weeks, while the relevant AI models only take a few hours. OpenAI said that the incident stemmed from the fact that when testing the AI's cybersecurity capabilities, some models broke through the “sandbox” environment that originally operated in isolation and entered the Internet environment. The models involved include GPT-5.6 Sol and two models that have not been publicly released. They discovered and exploited multiple vulnerabilities, which eventually led to the Hugging Face system being broken. OpenAI said it has launched an investigation with Hugging Face and notified the US government, law enforcement agencies, etc. of the relevant situation.

OpenAI's artificial intelligence model for cybersecurity testing unexpectedly broke through the AI platform Hugging Face's internal system last week and completed a series of vulnerability discovery and utilization operations. According to reports, if such attacks are carried out by human hackers, it can usually take weeks, while the relevant AI models only take a few hours. OpenAI said that the incident stemmed from the fact that when testing the AI's cybersecurity capabilities, some models broke through the “sandbox” environment that originally operated in isolation and entered the Internet environment. The models involved include GPT-5.6 Sol and two models that have not been publicly released. They discovered and exploited multiple vulnerabilities, which eventually led to the Hugging Face system being broken. OpenAI said it has launched an investigation with Hugging Face and notified the US government, law enforcement agencies, etc. of the relevant situation.