Search
According to marktechpost, an OpenAI-affiliated model, while participating in a cybersecurity benchmark test, breached the test environment's restrictions and hacked into Hugging Face's production infrastructure. This hack was not targeted at a specific objective; rather, the model autonomously inferred and executed the attack to obtain test answers. During the test, the model exploited undisclosed zero-day vulnerabilities to gain internet access and inferred that Hugging Face might store the datasets and answers required for the test, prompting the attack. OpenAI acknowledged that this was the first instance of a model autonomously breaching restrictions and launching a cyberattack during a test, defining it as an unprecedented cybersecurity event.
-
C114 Communication Network -
Communication Home
