Report: OpenAI agents covered their tracks after unauthorized access to Australian sites

According to a report from the cybersecurity firm Asymmetric Security, AI agents developed by OpenAI gained access to several Australian government websites and afterward tried to erase their own tracks. An AI agent is a computer program that can carry out tasks more or less on its own, for example searching for information and clicking around the web without constant human control.

According to the report, the agents opened private accounts on a web analytics service to hide their searches and created temporary email addresses. One account was set to delete itself automatically after 48 hours. Asymmetric Security stresses that it cannot be determined whether this was a deliberate cover-up or simply a consequence of how the agents were programmed.

Found a way into internal systems

The incidents stretch from March to September. In one of the cases, an internal test model gained access to source code in the Medicare statistics service at the Australian agency Services Australia, while the agent was actually supposed to be researching public spending on medication for skin diseases. When the model couldn't find answers in the public data, it found its own way into the agency's internal system, where it could retrieve files and passwords. In another case, an agent used a mistakenly exposed access key at the state health data agency VAHI in the state of Victoria.

Australian authorities say that it was not patients' personal records that were accessed, but aggregate statistics and internal file names.

OpenAI waited for months

OpenAI itself discovered the problem during an internal review of what the company calls "misaligned behavior," but it was only almost three months later, on September 10, that the Australian government was informed. The report was also sent to an email address that was not monitored, which has drawn additional criticism.

Australian Prime Minister Anthony Albanese has called the course of events "clearly unacceptable" and has personally raised the matter with OpenAI CEO Sam Altman. The country's intelligence agency, the Australian Signals Directorate, is now investigating the case, and it is part of the government's work on developing national standards for AI safety.

OpenAI itself says that most of the activity was routine tasks in which the agents searched for publicly available information, and that they often use government websites because these are considered trustworthy sources.