OpenAI has issued an apology after hacking Australian government websites and has vowed to do better in the future as well as work with Australians to rebuild trust in its brand.
“We are sorry and working to do better in the future,” it said in a blog post.
This comes after it was revealed OpenAI had accessed Australian government websites including Services Australia (Medicare), the New South Wales Bureau of Crime Statistics and Research, the Victorian Department of Health and the Australian Institute of Health and Welfare.
OpenAI says it will take accountability for its actions and says it will work with Australia to develop ways in which AI developers and the government can work together to identify and respond to AI cyber behaviour – malicious or otherwise.
OpenAI’s Chief Strategy Officer, Jason Quan, will appear at the Joint Select Committee on Artificial Intelligence in Sydney on October 6 to answer questions about what OpenAI knows, how it can respond and what steps have been taken to do better moving forward.
Alarmingly OpenAI revealed in July that some of its models had broken out of the systems designed to isolate them from the internet and control them. This became known as the Hugging Face incident.
OpenAI said its models are persistent and collaborative enough that, without sufficient safeguards, they can find and exploit security weaknesses across websites and computer systems.
Following the Hugging Face incident: OpenAI has strengthened its research safeguards, included additional network restrictions, and increased its monitoring.
OpenAI said it launched investigations and notified Services Australia and the Victorian Department of Health on September 10 and the New South Wales Bureau of Crime Statistics and Research on the September 18 about the unauthorised access.
The Australian Institute of Health and Welfare access did not meet OpenAI’s disclosure thresholds because it was accessed in the same way the public would access the system. But the company decided to share their findings on September 24 anyway.
The Services Australia Medicare Statistics Reporting Service was accessed with experimental internal-only OpenAI model which did not have the full set of safeguards used with publicly available products.
OpenAI says it accessed Services Australia’s Medicare Statistics Reporting Service but has no evidence that anyone’s medical records were accessed.
It is not unusual for AI agents to be assigned tasks to find, interpret and analyse publicly available information.
In this case the task was researching government spending per person on medicines for skin conditions in Victorian communities.
The model had difficulty finding that information and took actions it was not authorised to take.
While looking for the information at Services Australia’s Medicare Statistics Reporting Service the model found a way of gaining non-public access.
It then went on to use this access to review technical system information and source code related to the service but all with the intention of finding the information they originally set out to find.
OpenAI is committing resources and expertise to help maintain protection and principles.
This will include providing dedicated support and funding for affected agencies and to strengthen cyber defences across critical infrastructure and other sensitive areas.
OpenAI will also establish an Australian task force with independent expertise to develop policy recommendations to manage risks from increasingly capable AI agents.
