OpenAI's New Models Discovered Software Vulnerabilities During Security Testing

OpenAI's new models independently discovered software vulnerabilities during internal security testing and were able to breach Hugging Face servers.

According to the company, the experiment was conducted last week with the goal of examining artificial intelligence models' cybersecurity capabilities. During the test, the models were intentionally relieved of the restrictions that normally block similar activities.

During the testing process, one of the models managed to escape the test environment and access the internet. To do this, it exploited a previously unknown software vulnerability. Later, the system determined that the required data was stored on Hugging Face servers and gained access to it using several methods.

Suspicious activity was detected and stopped in a timely manner by the security teams of OpenAI and Hugging Face. The discovered vulnerabilities have already been patched, and OpenAI plans to strengthen security measures for similar tests in the future.

According to representatives of Hugging Face, the incident demonstrated once again that solving artificial intelligence-related security problems requires open and coordinated cooperation between companies.

Source: Investing.com