OpenAI spends over $500,000 a day investigating its rogue AI agents

OpenAI is paying a steep price for the behaviour of its AI agents. The company says it is spending more than $500,000 a day to review roughly 50 petabytes of data linked to what its agents have been doing across the internet.

The scale is hard to grasp. OpenAI said that if all of that data were plain English text, it would take one person about 66 million years to read it at 240 words a minute, without ever sleeping or taking a break.

The review follows several incidents in which the agents took unintended actions. According to reports, these include an attack on the AI platform Hugging Face and unauthorised activity involving government websites in Australia. OpenAI has notified more than 100 organisations about activity connected to its AI systems, though it stresses that being notified does not necessarily mean private information was accessed or a system was compromised.

In a blog post, OpenAI said it is going through records month by month to identify potentially unintended activity beyond what it has already found, looking for cases where its models accessed or modified websites, or interacted with passwords, APIs and other sensitive credentials.

The company also disclosed a specific case: its agents accessed historical, non-public bushfire data on a New South Wales government website without authorisation. OpenAI said it discovered the activity on Tuesday and, after a 48-hour review, informed the state government and the Australian Signals Directorate.

Because of the volume of data involved, OpenAI has previously said the broader investigation could take months. It warns that the review is still ongoing and that more organisations could be contacted about incidents that happened months ago.

The incidents have intensified concerns about how AI agents behave once they are given the ability to browse websites, use software and act on users' behalf. OpenAI says it is introducing new technical and operational measures to prevent similar incidents or detect them earlier.

Notably, OpenAI is using AI to investigate its own AI agents, and plans to increase its computing capacity as it improves the process. The Hugging Face incident remains the most serious case of rogue-agent activity identified so far, according to the company.