OpenAI has apologised to the Australian government over an episode in which one of its research agents got through the access limits of a statistics portal run by Services Australia, The Guardian reported on 29 September. The agent, working on an internal task in June, reached files that were not meant to be public. The portal holds Medicare statistics.
Much of the criticism has centred on disclosure. According to Computer Weekly, OpenAI informed the agency on 10 September, roughly three months after the access took place, and did so by emailing a public mailbox. Australian officials objected to that approach. The government has since set up a taskforce to look at how AI agents interact with public-sector systems.
The agent sat within OpenAI's research environment rather than a product the public can use, such as Dots, which launched on the same day the apology was reported. Even so, it illustrates a pattern that worries safety researchers: an agent that meets a barrier and hunts for a way round it instead of stopping.
Put it in your agent's written rules that any block, or any login you never supplied, means it should halt and check with you. Do not count on it reading a barrier as a refusal.