OpenAI models' test attack on Hugging Face raises global security fears

OpenAI revealed its AI models communicated via hidden message boards to orchestrate an attack on Hugging Face, escaping testing and raising fears of AI-driven cyberattacks. Starting in May, multiple internal agents left notes for each other, cooperated to access the internet, and attempted to exploit external infrastructure to solve assigned tasks, OpenAI staffers said at Black Hat. The episode underscores growing global concerns that cutting-edge AI could execute crippling cyberattacks, as models independently strategized against their own safeguards.