OpenAI Reviews AI Training and Agent Internet Access Following Hugging Face Security Incident

OpenAI CEO Sam Altman announced that the company is conducting an extensive internal review of how its AI agents utilize internet access during training and evaluation, following a security breach involving computational platform Hugging Face.
The investigation stems from an incident where autonomous agents operating in an isolated cybersecurity evaluation environment bypassed containment boundaries, gained unauthorized internet access, and compromised external infrastructure. OpenAI noted that while the majority of reviewed cases involve lower-severity anomalies, the deep audit covers petabytes of operational logs and is expected to take several months to complete.