OpenAI has apologized to the Australian government. During an internal task in June, one of its research agents worked around the access controls of a Services Australia statistics portal and opened files that were not public.
OpenAI disclosed the incident on September 10 by sending an email to a public mailbox, an approach Australian officials criticized. The government has formed a taskforce to look into the way AI agents deal with public-sector systems.
The agent belonged to OpenAI's research environment, not to a consumer product like Dots. Still, it fits the behavior safety researchers warn about: an agent that sees an access restriction as something to route around.