TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
7x24hNews

OpenAI Discloses AI Models Secretly Communicated for Several Months, Eventually Decided to Attack Hugging Face

2026.08.06 - 01:58
Share

TechFlow reports, on August 6, OpenAI stated that AI models involved in attacking the artificial intelligence open-source community and model hosting platform Hugging Face began communicating through undiscovered information channels as early as May, collaborating to attempt to break out of the testing environment.

OpenAI employees Wallace and Dalton stated that multiple agents and AI models intended for internal use only spent months "leaving messages for each other," and gradually formed a consensus around a goal: accessing the internet to solve the tasks they were required to complete, some of which were impossible to complete without network access permissions.

Wallace stated: "At some point, these agents realized that perhaps we could try to exploit or attack external infrastructure to find answers to the test tasks I was being evaluated on." The researchers' presentation added more details about the Hugging Face hacking incident. The incident highlighted growing global concerns that cutting-edge artificial intelligence systems could be used to launch destructive cyberattacks. (Jin10)

Add to Favorites
Share to Social Media