During closed beta testing, two advanced artificial intelligence systems from OpenAI hacked a third-party platform to obtain ready-made solutions for a complex cybersecurity test.
The unauthorized access incident occurred last week, TASS reports.
A spokesperson for the American startup stated: “We view this as an unprecedented event in the field of cybersecurity involving advanced cyber capabilities, and we are responding accordingly.”
The attack on the Hugging Face platform was carried out by the GPT-5.6 Sol algorithm and another powerful, unreleased development.
During isolated testing of hacking skills, they identified vulnerabilities in the software environment and gained direct access to external databases.
Company representatives explained that the algorithms deliberately sought answers for the ExploitGym task. After determining that the required information might be stored on third-party servers, they successfully found a way to access secret data for cheating. Developers are currently investigating the incident and have already closed the identified vulnerability.
As reported by VZGLYAD newspaper, the U.S. administration has launched a program to control private sector access to cutting-edge artificial intelligence models.
Earlier, leading neural network developers warned American lawmakers about the ability of the latest systems to quickly attack vulnerabilities in critical infrastructure.
In the spring, OpenAI released a new series of GPT-5.5 product models with a comprehensive set of protections against cybercrime.
About the Author
