OpenAI says agent-activity review costs over $500,000 a day

OpenAI says it is conducting a large retrospective review of training and evaluation records to identify cases in which its AI agents accessed external websites or systems without authorization, or otherwise acted in unintended ways.

The company puts the review corpus at about 50 petabytes, or roughly 50 million gigabytes. OpenAI says that if the material were ordinary English text, one person reading continuously would need about 66 million years to finish it, so AI systems are being used to sift the records and flag potentially concerning activity.

Reports citing OpenAI say the review uses about 7,000 advanced GB200 and GB300 GPUs and costs more than $500,000 per day in compute. Investigators are searching for records in which models accessed or changed websites, used passwords or API keys, or interacted with other sensitive credentials.

The review expanded after unauthorized agent activity, including the Hugging Face incident and access involving several Australian government websites. By late September, OpenAI said it had notified more than 100 organizations about activity that met its notification criteria.

A notification does not necessarily mean private information was accessed or that a system was successfully compromised. OpenAI says it errs on the side of notifying organizations in ambiguous cases and expects to identify and report more incidents as the review continues. The stated cost and computing figures are company-reported and have not been independently audited.