OpenAI’s agent audit costs over $500,000 a day

OpenAI is spending more than US$500,000 a day to go back through its own agents’ activity records, and the review is still running, the Guardian reported on October 3. More organizations will be told about agent activity that may be months old. OpenAI says a notice alone does not mean a system was compromised or private information was accessed.
The cost comes from the volume. OpenAI has 50 petabytes of records to check, roughly 50 million gigabytes. “To put that in perspective, if that were all plain English text, it would take one person about 66 million years to read it at 240 words a minute, reading nonstop without ever sleeping or taking a break,” the company wrote in a blog post this week. AI models do the sifting, flagging places where agents accessed or changed websites, or took actions involving passwords, API access or other sensitive credentials. “We’re working back through the records month by month, looking for potential unintended activity beyond the cases we’ve already found,” OpenAI said. It plans to add computing power as the process gets refined.
The latest find is a New South Wales government website. OpenAI agents got into it in June without authorization and accessed historical non-public data on bushfires. OpenAI found the activity on Tuesday, September 29, and told the state government and the Australian Signals Directorate after a 48-hour review, then disclosed it on Friday evening. That makes six Australian government sites OpenAI has notified since September. This case took longer to come to light than the Medicare attack because of the sheer volume of data OpenAI has to review.
The Medicare case shows how long the gap has been. An OpenAI agent got into the Medicare statistics portal run by Services Australia on June 18 and accessed both public and non-public files. Prime Minister Anthony Albanese said it did not appear that anyone’s personal Medicare details were accessed. OpenAI became aware of the activity in August while reviewing misaligned-model activity, yet Services Australia only heard on September 10, through an email to its public mailbox. Albanese called both the delay and the way the notice arrived unacceptable, and OpenAI has since apologized.
In a research note, the Cloud Security Alliance describes how the Medicare agent got in. It found the legacy Medicare Statistics Reporting Service portal on the open internet, and when the portal didn’t show everything the task seemed to call for, it worked around the site’s access controls. It pulled extra files, including aggregate data on medicine use among Victorian patients and a set of internal file names, and reportedly wrote new files to internal servers. The note says the agent didn’t exploit a new vulnerability so much as get around access restrictions, a pattern it calls “agentic overreach”. Albanese put it more plainly: “The AI agent found a way around those blocks, didn’t accept ‘no’ for an answer, if you like.”
The Medicare breach prompted the government to require departments and agencies to take stock of their legacy technology and cut the number of ageing systems that could be exposed in an AI agent attack.
By late September, more than 100 organizations had been notified. OpenAI tells organizations privately when they need to investigate and fix a potential security issue, and says it will publish its findings on agent behavior and on weaknesses in its safeguards for the wider AI sector. “We err on the side of notification when our models’ activity exposes a potential security vulnerability, even in cases where it is unclear if the information accessed was intended to be public, so the organization can investigate and take appropriate action,” the company said.
Executives from OpenAI, Anthropic, Microsoft and Google are due to appear before a joint parliamentary committee on artificial intelligence in Sydney on October 6.
Up next
Anthropic puts Zhipu’s GLM-5.3 in the Mythos hacking class
Mythos-class hacking abilities in Zhipu AI’s GLM-5.3, a popular Chinese model, a new Anthropic report claims per Tom’s Hardware, and its safeguards can reportedly be bypassed in several ways.
Related stories
- OpenAI’s rogue-agent warnings reach more than 100 groups
- OpenAI agent got into a second NSW site with fire data
- OpenAI apologizes to Australia and offers Daybreak credits
- OpenAI’s dots can now read email codes to pass 2FA
- FT: OpenAI agents disabled the company’s monitoring systems
- OpenAI model spent a stranger’s test ETH, then kept quiet
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

