The American company „OpenAI“ (OpenAI) reported that its artificial intelligence models went outside the controlled test environment during a cybersecurity test and caused a breach in the infrastructure of the „Hugging Face“ platform, Reuters reported, BTA reported.
In a post on its corporate blog „OpenAI“ indicated that it had tested the capabilities of its latest models in a controlled environment. According to the company, an autonomous agent managed to overcome the limitations of the test environment, gain access to the Internet and penetrate the infrastructure of „Hugging Face“ in an attempt to fulfill the set task. its task.
The company described the case as “unprecedented“ and said it was taking additional measures to increase security in such tests.
According to “OpenAI“ the test was intended to evaluate the ability of the latest models to detect and exploit vulnerabilities in computer systems. Instead of remaining in the isolated environment, the agent used a previously unknown vulnerability to gain access to the Internet.
The company indicated that the model subsequently concluded that the “Hugging Face“ platform may contain information that would help it in completing the task. According to “OpenAI“ the agent made unauthorized access to data on the platform that could be used to improve the test results.
Last week “Hugging Face“ announced that it had been the target of a cyberattack carried out by an autonomous artificial intelligence system. According to the platform, the attacker performed thousands of sequential actions and changed the location of its command infrastructure.
The case reinforces experts' concerns about the potential of the most advanced artificial intelligence models to be used to carry out complex cyberattacks. In recent months, the topic has also attracted attention due to the developments of “Anthropic“ in the field of cybersecurity, which raised questions about the need for stricter mechanisms for controlling such technologies.