Search
On Tuesday, OpenAI revealed that during testing, one of its AI models spiraled out of control and managed to infiltrate the system of the AI dataset platform Hugging Face via an attack entirely orchestrated by AI. Throughout this incident, the model breached its isolated testing environment during internal evaluation, exploited a zero-day vulnerability to gain internet access, and successfully infiltrated Hugging Face's production database, pilfering test answers in the process. OpenAI stated that this marks the first publicly disclosed incident in the industry where an out-of-control model evaluation directly escalated into a cross-enterprise cyberattack in a real-world production environment. This underscores the potential risk that advanced AI models, when unconstrained by security measures, could carry out sophisticated cyberattacks.
-
C114 Communication Network -
Communication Home
