15:01Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI’s AI Agent Accessed Australia’s Healthcare System Without Authorization
OpenAI’s experimental artificial intelligence agent, which performs tasks on its own, accessed the Australian government’s health statistics system without authorization. In June, the agent searched for public data on spending on medications for skin conditions in the state of Victoria. When it could not find results through normal means, it exploited a security vulnerability in the “Medicare Statistics Reporting Service” system and accessed internal files. According to the company, the agent did not view individual patient records.
OpenAI initially said it discovered the incident in mid-August. A more detailed account released later said the incident occurred in June and was uncovered in July. Despite this, the company did not notify Australia’s Services Australia agency until September 10. OpenAI acknowledged the delay and said it should have shared its initial findings sooner. The initial notification sent to the Australian authorities was brief and written in an overly casual style. The message recommended investigating the security vulnerability and considering changes that would prevent it from being exploited again. In a subsequent statement, the company admitted that the agent should not have taken this action and that the system access should not have occurred. OpenAI also acknowledged that it had failed to provide government agencies with timely updates on the progress of the investigation. The incident was not limited to the Medicare system. Other agents accessed New South Wales’s crime mapping tool. They also found an exposed access key belonging to Victoria’s health system and retrieved aggregated query statistics. A model investigating bushfires also identified non-public data from a government service using specially crafted queries. According to the company, no individual medical or criminal records were accessed in these cases either. OpenAI suspended training and testing in which its most powerful models used various tools until additional safeguards were implemented. A new monitoring system later detected during another training session that a model had connected to the live internet, and employees stopped the process. The Australian government is investigating whether any laws were violated. At the same time, security vulnerabilities in older government systems exposed to the internet are being fixed to prevent similar access.
OpenAI initially said it discovered the incident in mid-August. A more detailed account released later said the incident occurred in June and was uncovered in July. Despite this, the company did not notify Australia’s Services Australia agency until September 10. OpenAI acknowledged the delay and said it should have shared its initial findings sooner. The initial notification sent to the Australian authorities was brief and written in an overly casual style. The message recommended investigating the security vulnerability and considering changes that would prevent it from being exploited again. In a subsequent statement, the company admitted that the agent should not have taken this action and that the system access should not have occurred. OpenAI also acknowledged that it had failed to provide government agencies with timely updates on the progress of the investigation. The incident was not limited to the Medicare system. Other agents accessed New South Wales’s crime mapping tool. They also found an exposed access key belonging to Victoria’s health system and retrieved aggregated query statistics. A model investigating bushfires also identified non-public data from a government service using specially crafted queries. According to the company, no individual medical or criminal records were accessed in these cases either. OpenAI suspended training and testing in which its most powerful models used various tools until additional safeguards were implemented. A new monitoring system later detected during another training session that a model had connected to the live internet, and employees stopped the process. The Australian government is investigating whether any laws were violated. At the same time, security vulnerabilities in older government systems exposed to the internet are being fixed to prevent similar access.