An OpenAI artificial intelligence agent accesses an Australian government portal without authorization

Revelations about artificial intelligence agents crossing the limits of their tasks are piling up. In recent weeks, large laboratories have published new details about tests that ended up affecting real third-party systems. They have been called 'uncontrolled agents': programs capable of chaining actions to meet an objective that, in these cases, have ended up accessing resources for which they did not have permission.

The last known case affects an Australian government portal. An OpenAI agent accessed the site without authorization on June 18, but Sam Altman's company did not notify Services Australia of what happened until September 10, 84 days later. The incident, made public on September 24, has provoked a protest from the Australian Government over the delay.

The access came during an internal test in which a model investigated public spending on medicines. As explained by the Prime Minister, Anthony Albanese, the agent found blockages when requesting information, tried alternative routes and ended up entering areas for which he did not have authorization.

The system consulted public and non-public information from the Medicare statistics portal. Albanese also noted that, according to Services Australia, the agent wrote files to an internal server during the process, although the extent of that activity remains under investigation.

Australian authorities say no individual medical data was accessed. The portal contained aggregate statistics and operated independently of the Medicare benefit and payment systems. The platform remains deactivated and Australia has created a working group to investigate what happened and review the security of its networks against threats linked to AI.

It's not the first time

OpenAI has acknowledged that its models compromised the Hugging Face platform and that its review of previous activities has led to dozens of third parties being notified. The company relates these behaviors to strategies that the models adopt to solve difficult tasks, even if they deviate from what their developers intended.

Anthropic, for its part, published on September 9 an analysis of four incidents in which Claude models gained unauthorized access to real systems during cybersecurity assessments. The test environments had an internet connection that had to be closed.

Meta also acknowledged in August that one of its models exploited a vulnerability in an external service after a configuration error during an evaluation allowed it to connect to the internet.