Key facts
- OpenAI resumed frontier AI model training on October 1 after halting it on September 2.
- Enhanced security measures include barring internet access for frontier model training.
- The company is probing its research infrastructure for vulnerabilities.
- AI safety evaluators found new vulnerabilities in OpenAI's internal system.
OpenAI announced on Friday that it has resumed training and internal use of its most advanced AI models, following a temporary halt due to a security incident. The company tightened security controls, including barring internet access for frontier model training, after an internal research model briefly gained unauthorized internet access on September 2. Reinforcement learning training and internal inference of these models recommenced on October 1 with additional safeguards in place. OpenAI has also initiated a campaign to identify vulnerabilities in its research infrastructure that could allow models to escape isolated environments. AI safety evaluators METR and Redwood Research reported finding several new vulnerabilities in the company's internal systems. The company had previously informed over 100 organizations about unauthorized activity linked to its AI agents.