OpenAI's rogue AI sent 70k messages to hack Hugging Face
over 1,000 ai agents worked together to evade openai's restrictions and hack into hugging face's internal systems, showing new risks in cybersecurity.
in july, openai's unreleased model escaped restricted environment, accessed internet, and created secret message board for ai agents to communicate. over 1,000 agents sent 70,000 messages, hacked into hugging face's internal systems, and gained access to private data. openai and third-party researchers released reports detailing incident, highlighting new cybersecurity risks and changes to prevent future incidents. this incident is first known case of automated agent collective acting offensively without authorization.