OpenAI is currently spending more than $500,000 (€444,000) a day on investigating security breaches involving its AI agents. This includes research into unauthorised intrusions into various Australian government platforms. The company is analysing an enormous dataset of 50 petabytes (50 million gigabytes) to detect unintended activities. To process this overwhelming volume of information, the company uses artificial intelligence (AI) to scan for instances where its models have modified websites or gained access to sensitive login credentials and APIs.
Recent findings show that OpenAI agents illegally gained access in June to historical, non-public data on bushfires on a New South Wales government website. This is the sixth Australian government body to be alerted to such activities since last month, following an earlier breach of the Medicare statistics portal managed by Services Australia. The delay in reporting the incident in New South Wales was attributed to the enormous amount of data that had to be thoroughly examined.
To date, OpenAI has contacted more than 100 organisations in connection with possible attacks. However, the company clarified that these notifications do not necessarily confirm that a system has been compromised or that private data has been stolen. Instead, OpenAI gives priority to transparency and alerts organisations whenever a potential vulnerability is detected, so they can take the necessary security measures. The company plans to make these findings public to help the wider AI sector improve its security measures.
Australia’s response
Following the Medicare incident, the Australian government has instructed its agencies to review legacy technology. The aim is to minimise the cybersecurity risks associated with outdated systems. In the meantime, the investigation is still ongoing, while OpenAI continues to work through the data month by month. As part of these developments, representatives from Google, Microsoft, Anthropic and OpenAI will appear in Sydney next Tuesday before a joint parliamentary committee on AI.
Business AM
Ltd
