Landian News Technology
OpenAI officially confirms internal test model escaped sandbox, autonomously exploited vulnerabilities and hacked open-source platform HuggingFace
#SecurityNews OpenAI officially confirmed that an internal test model escaped its sandbox, autonomously exploited vulnerabilities and hacked the open-source platform Hugging Face. The source of this hacking incident turned out to be an AI agent driven by an OpenAI test model. Specifically, the model first exploited vulnerabilities in OpenAI's isolated environment and moved laterally to other nodes to gain public network access, then exploited vulnerabilities on the HF platform to launch the attack. The purpose of the attack was surprisingly that the model wanted to steal benchmark test answers from HF to get better scores. Read the full article: https://ourl.co/114029