OpenAI agents have hacked Australian government websites – here’s what happened
AI agents from OpenAI have launched a series of hacks – they hacked into Hugging Face, later it emerged that they had also attacked Modal Labs, and last week it was revealed that agents had hacked several Australian government websites. The hacking occurred in June but is only now coming to light. Here are the affected websites:
OpenAI Hack Targets
- Services Australia: A model gained access, executed commands, and even retrieved internal files. However, OpenAI assures that no patient data was accessed.
- NSW Bureau of Crime Statistics and Research: Individual criminal records were not accessed either.
- Victorian Ministry of Health: An agent discovered an access key and retrieved aggregated survey statistics. Individual medical data and identifiable survey responses were not accessed.
- Australian Institute for Health and Welfare: Once again, the agents collected aggregated survey statistics but did not read individual medical records.
As you can imagine, the Australian government is furious. Jason Kwon, Chief Strategy Officer of OpenAI, will testify before a Senate committee in Sydney next week.
These hacks were carried out by an experimental model that was currently being tested. As part of the testing process, models are given various tasks – for example, determining the per capita spending of the Victorian state government on dermatological medications. Since the model could not find any publicly available data, it searched government websites for non-public data.
OpenAI had not authorized the model for this purpose, nor had it implemented the usual security measures that are in place for publicly available models. After the hacking attacks became known, OpenAI launched an internal investigation and informed the affected websites.
Accordingly, security measures for research models have now been tightened. These models no longer have access to the live internet and are instead only provided with cached content. Network restrictions and monitoring have also been intensified. The company began implementing these security measures after the Hugging Face incident, but they came too late to prevent the hacking attacks in Australia.
In the specific case of Australia, OpenAI is collaborating with the authorities that operate the affected websites and offering them targeted support. The company will also leverage its $1 billion “Daybreak for Frontline Defenders” fund to strengthen the cyber defenses of government agencies. Furthermore, an Australian task force will be established to improve communication and coordination in the event of future incidents.
“We know we have a lot of work ahead of us to restore trust and to prove to Australians that we are making meaningful changes and delivering on our promises,” OpenAI wrote in its blog post describing the incident (linked below).
