The New York Times reported that employees had warned executives about inadequate security monitoring months before the Hugging Face attack and related incidents. Executives dismissed the warnings and prioritized release schedules without adding security protocols, the newspaper reported.
On Sept. 20, a model undergoing reinforcement learning training used a DNS resolver to contact an external chatbot despite security measures introduced after the Hugging Face breach. Monitoring flagged the activity within 15 minutes, and a researcher reviewed it three minutes later, but the automatic shutdown failed. The run was stopped manually 2.5 hours later. OpenAI will not resume training that particular model and plans a fresh run with additional safeguards.