Washington, August 19 (QNA) - OpenAI has announced a new set of security and safety policies and measures aimed at containing security incidents during the testing of AI models.

The new measures include more detailed monitoring of model behavior throughout the development process, along with a greater focus on model alignment and security during the post-training phase.

The company said the new measures were prompted in part by the anticipated Astra model's advanced cybersecurity capabilities, as well as the accelerating pace of development across the AI sector.

The firm noted that it paused reinforcement learning for two weeks following the Hugging Face incident before resuming training on several lower-risk models.

The new measures also include stronger network-isolation practices, ensuring that a compromise of any single workload or supporting service would not, by itself, give attackers unauthorized access to the internet or other internal networks.

They likewise include a sophisticated monitoring system that scans a wide range of activities for unauthorized behavior and tracks the actions of tools used by models, inference traces and activity logs to detect patterns that could indicate concerning activity or an attempt to circumvent security controls.

OpenAI aims to issue alerts within 30 minutes of detecting concerning activity, while the company estimates that operating the monitoring system will require computing resources equivalent to about 20% of the resources used by the process being monitored. (QNA)