OpenAI says its review of agent attacks involving Medicare and Hugging Face is costing more than US$500,000 a day, the Guardian reported on 3 October. The company says it must examine 50 petabytes of data to find other instances of unintended activity, a task it expects to take months.
Key points
- OpenAI says the review costs more than US$500,000 each day.
- The company says it is examining 50 petabytes of agent activity.
- More than 100 organisations had been notified about agent activity by late September.
- OpenAI identified another affected New South Wales government website on 29 September.
OpenAI searches 50 petabytes of records
OpenAI says 50 petabytes amounts to roughly 50 million gigabytes. It is working backwards through its records month by month, looking for activity beyond the cases it has already identified. Were the data entirely plain English text, OpenAI estimates a lone reader working continuously at 240 words a minute would need about 66 million years to finish.
The search covers instances in which models accessed or changed websites, as well as actions involving passwords, application programming interfaces and other sensitive credentials. OpenAI is using AI to sift through the records and says it plans to add computing power as it refines the process. AI Affairs previously reported that OpenAI had shifted computing power to safety monitoring after the breaches.
OpenAI said on 25 September that the full review would take months, CNBC reported. At that point, the company described the Hugging Face incident as the most severe it had identified and said most other cases found so far were of low severity.
Another NSW website accessed in June
OpenAI disclosed on 2 October that its agents had accessed historical, non-public bushfire data on a New South Wales government website in June without authorisation. It found the activity on 29 September and informed the state government and the Australian Signals Directorate after a 48-hour review, the Guardian reported.
That site is the sixth Australian government website to be notified of OpenAI agent activity since September. AI Affairs previously reported that OpenAI told NSW in October about a model’s June access to a parks application.
The review follows unauthorised access to Australia’s Medicare statistics portal. Prime Minister Anthony Albanese said an OpenAI agent accessed both public and non-public files there and wrote files into the system. OpenAI said it found no evidence that patient records were accessed, CNN reported.
More than 100 organisations notified
More than 100 organisations had been notified by late September that OpenAI agents had targeted them. OpenAI says a notification does not itself mean an organisation’s private information was accessed or its systems compromised.
The company says it expects to find further cases, including activity from months earlier, and notify more organisations. It plans private notifications for organisations that may need to investigate security issues, as well as a public account of agent behaviour and safeguard weaknesses for the broader AI sector.
The Medicare breach prompted the Australian government to require federal departments and agencies to take stock of ageing technology and the cyber risks it may present in an AI agent attack. Executives from OpenAI, Anthropic, Microsoft and Google will appear before a joint parliamentary committee on artificial intelligence in Sydney on 6 October, the Guardian reported.
