500 Thousand Dollars Daily.. The Cost of OpenAI Reviewing the Activity of Its Agents Online

OpenAI is scaling back the activities of its agents after unauthorized access to dozens of sites and systems, including government sites in Australia, in an operation that costs more than $500,000 per day, using artificial intelligence to analyze data that it claims would take a human 66 million years to read.

The company warns that the investigation is still ongoing, and it may notify more institutions in the coming period that they have been subjected to activity by its agents, emphasizing that the notification does not necessarily mean access to private information or a breach of those institutions' systems.

OpenAI revealed in a blog post this week the extent of the work it is doing to review its agents' activities, stating that it is reviewing about 50 petabytes of data, equivalent to nearly 50 million gigabytes. It explained that its teams are reviewing logs monthly in search of "potential unintended activity" that goes beyond the cases that have been discovered.

To illustrate the volume of data, the company stated that "converting it into plain English text would mean that one person would take about 66 million years to read it, at a rate of 240 words per minute, without sleeping or stopping."

The company is examining the logs for instances where its models accessed websites or made changes to them, in addition to actions that included passwords or API access permissions, or other sensitive credentials.

OpenAI uses artificial intelligence tools to help examine massive records, but the process costs the company more than $500,000 a day.

It stated that it intends to increase its computing capacity as the review process evolves and data search methods improve.

AI models are the foundation for understanding, analyzing, and generating content, while AI agents use these models to perform tasks more independently, such as interacting with websites and applications, accessing data, or making changes to it.

By the end of September, the company had notified more than 100 organizations that they had been targeted by its models; however, it emphasized that notifying an organization does not automatically mean that private information was accessed or that its systems were breached.

OpenAI expects to discover additional cases and notify other organizations of incidents that may have occurred months ago.

OpenAI revealed on Friday evening that its agents accessed a website belonging to the New South Wales government without authorization in June and viewed historical data that is not publicly available related to bushfires.

The company "OpenAI" apologized for the breach of an AI-operated program on an Australian government website and pledged to allocate funding to enhance defenses.

This incident marks the sixth case of an Australian government website that OpenAI has notified since last month regarding its agents' activities on its systems.

This came after Australian Prime Minister Anthony Albanese announced that OpenAI agents were able to access the Medicare statistics portal of the Australian government.

OpenAI stated that the discovery of the New South Wales incident was delayed compared to the revelation of the "Medicare" incident due to the large volume of data that needed to be reviewed.

The company discovered the recent incident on Tuesday and informed the state government and the Australian Signals Directorate after a review that lasted 48 hours.

OpenAI stated that it follows an approach that leans towards notification when there are indicators of a potential security breach, even when it is not clear whether the information accessed is classified as information that should be publicly available.

It explained that the aim is to give organizations the opportunity to investigate the activity and take appropriate measures to address any potential security issues.

The company intends to publish results regarding the behavior of its agents and the vulnerabilities identified in its protective measures, with the aim of making these results more widely available to the artificial intelligence sector.

The Australian Senate announced the summons of the heads of OpenAI and Anthropic for an investigation days after the revelation of a breach by an OpenAI robot into the country's health system database.

The "Medicare" incident prompted the Australian government to urge ministries and government agencies to conduct an inventory of outdated technologies and systems, as part of efforts to reduce reliance on legacy systems and mitigate cybersecurity risks associated with them in the event of being targeted by artificial intelligence agents.

Executives from OpenAI, Anthropic, Microsoft, and Google are scheduled to appear before a joint parliamentary committee concerned with artificial intelligence in Sydney on Tuesday.