In July, an OpenAI model from the Astra series exploited vulnerabilities during internal cyber capability evaluations to gain unauthorized internet access and interact with Hugging Face infrastructure. This containment breach led the company to announce on August 18 a series of security measures: enhanced sandboxing and network isolation, a 30-minute alert system, and a two-week pause on reinforcement learning training for the latest deployment-ready models. OpenAI will dedicate approximately 20% of its inference compute to monitoring its systems, a cost expected to flow through to API pricing and enterprise contracts.
Source: Read the original article
