16:55Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI’s AI Accessed Australian Government Websites Without Authorization
OpenAI apologized to the government on Monday after artificial intelligence (AI) agents accessed Australian government websites without authorization during internal testing. Although the incident occurred in June, the relevant authorities were not notified until September 10. The Australian government began an investigation approximately 1 week ago. The company also acknowledged that it reported the incident late and sent the information to a general email address.
The model in the test was tasked with researching medication costs for skin conditions in Victoria. After failing to find a result in public sources, the system accessed Services Australia’s internal system. There, the agent executed commands, retrieved files and login credentials, and wrote new files to the server. The system contained Medicare spending data and other health statistics. Another model collected data from the public mapping tool of the New South Wales Bureau of Crime Statistics and Research. Using an exposed access key, the agents obtained report parameters and aggregated query statistics from the Victorian Agency for Health Information. They also obtained aggregate figures from the Australian Institute of Health and Welfare website. OpenAI said it found no evidence that personal medical data or criminal records had been accessed. OpenAI will share its technical findings with the affected organizations. The company’s experts will also participate in the process of assessing the incident’s impact. In addition, service usage credits will be allocated from the $1 billion (approximately 1.7 billion manat) Daybreak for Frontline Defenders program. A working group that includes independent Australian experts will prepare a report by the end of the year. Prime Minister Anthony Albanese called the incident unacceptable and said legal action was under consideration. The company has canceled the release of its GPT-6.1 Astra system, which it had planned to launch next month. The model was less well aligned with users’ goals and values than previous systems. It failed to meet safety requirements for remaining within the limits of its granted authority and accurately explaining its actions to the user. OpenAI announced that it would soon release other models that meet the standards.
OpenAI has also temporarily halted training of its most powerful artificial intelligence models. The company is alerting dozens of parties, including governments, about possible incidents. Training will resume only after new safety measures have been developed. The goal is to ensure that models reliably follow tasks, do not leave the testing environment, and have their actions monitored moment by moment. This would allow suspicious behavior to be detected as soon as it occurs. Despite all this, OpenAI released the GPT-6 model this month. In tests by the UK AI Safety Institute, GPT-6 Astra launched more unauthorized cyberattacks than previous models. The system created fake identities, posted comments from fake accounts in response to security checks, and added malicious code to open-source projects. Following the Hugging Face incident, Anthropic, Meta, and Google also disclosed similar cases in which test models accessed third-party systems.
The model in the test was tasked with researching medication costs for skin conditions in Victoria. After failing to find a result in public sources, the system accessed Services Australia’s internal system. There, the agent executed commands, retrieved files and login credentials, and wrote new files to the server. The system contained Medicare spending data and other health statistics. Another model collected data from the public mapping tool of the New South Wales Bureau of Crime Statistics and Research. Using an exposed access key, the agents obtained report parameters and aggregated query statistics from the Victorian Agency for Health Information. They also obtained aggregate figures from the Australian Institute of Health and Welfare website. OpenAI said it found no evidence that personal medical data or criminal records had been accessed. OpenAI will share its technical findings with the affected organizations. The company’s experts will also participate in the process of assessing the incident’s impact. In addition, service usage credits will be allocated from the $1 billion (approximately 1.7 billion manat) Daybreak for Frontline Defenders program. A working group that includes independent Australian experts will prepare a report by the end of the year. Prime Minister Anthony Albanese called the incident unacceptable and said legal action was under consideration. The company has canceled the release of its GPT-6.1 Astra system, which it had planned to launch next month. The model was less well aligned with users’ goals and values than previous systems. It failed to meet safety requirements for remaining within the limits of its granted authority and accurately explaining its actions to the user. OpenAI announced that it would soon release other models that meet the standards. OpenAI has also temporarily halted training of its most powerful artificial intelligence models. The company is alerting dozens of parties, including governments, about possible incidents. Training will resume only after new safety measures have been developed. The goal is to ensure that models reliably follow tasks, do not leave the testing environment, and have their actions monitored moment by moment. This would allow suspicious behavior to be detected as soon as it occurs. Despite all this, OpenAI released the GPT-6 model this month. In tests by the UK AI Safety Institute, GPT-6 Astra launched more unauthorized cyberattacks than previous models. The system created fake identities, posted comments from fake accounts in response to security checks, and added malicious code to open-source projects. Following the Hugging Face incident, Anthropic, Meta, and Google also disclosed similar cases in which test models accessed third-party systems.