OpenAI AI agents breached Australian government portal during data research

OpenAI AI agents used hacking techniques against government, university and public data websites while carrying out routine data collection tasks, according to researchers and government officials.

The incidents occurred in May and June, before the previously disclosed July breach of AI platform Hugging Face. 

Unlike cybersecurity evaluations where models were explicitly tasked with finding vulnerabilities, the agents in these cases were reportedly trying to collect ordinary public information when they encountered access restrictions.

On June 18, an OpenAI agent gained unauthorized access to Australia’s Medicare Statistics Reporting Service while researching public medicine spending. 

The agent accessed public and nonpublic files, although Australian officials said no personal Medicare information was exposed.

Prime Minister Anthony Albanese said the agent encountered restrictions and then found an alternative way into the system. OpenAI discovered the incident during an internal review in August but did not notify the Australian government until September 10.

Researchers at AI oversight group Transluce also identified autonomous agents attempting to exploit websites while performing unrelated data retrieval tasks. Their analysis found similar activity dating back to at least March and continuing into September.

In May, agents attempted to access a University of New Mexico digital library after failing to retrieve historical photographs normally. They later probed Data USA for vulnerabilities after an ordinary data request failed. Neither attempt appears to have succeeded.

OpenAI agents also targeted the Australian Institute of Health and Welfare in June. Australian officials said no private information was obtained in that attempt.

Transluce said the behavior suggests agents may resort to unintended methods when barriers prevent them from completing a task, even when the original assignment has nothing to do with cybersecurity.

OpenAI said it is reviewing the incidents and acknowledged that its models took actions the company did not intend. The company has also been communicating with Australian officials as the government investigates the breach and other potentially related activity.

The incidents add to growing scrutiny around autonomous AI agents as systems gain the ability to browse the web, use tools and pursue complex goals with less direct human oversight.