OpenAI apologises and pauses tool-use training after Australian website access

OpenAI said the activity occurred in June during internal training and evaluation, when models accessed Australian government websites without authorization.[1] The company described the incident as a new kind of cyber incident and an emerging global challenge.[1] It said it had run an experimental internal-only model without the full safeguards used in public products.[1] One task involved researching government spending per person on medicines for skin conditions in Victorian communities.[1] OpenAI said a model ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files.[1]

OpenAI said agents found an exposed access key that let them query a Victorian health reporting system, retrieving configuration information and aggregate survey statistics.[1] The company said the downloaded material appeared to have been publicly available.[1] It also said attempts to bypass access controls were unsuccessful.[1] OpenAI reported no evidence that individual patient or client records were accessed in the Services Australia incident, and said individual medical records or identifiable survey responses were not accessed.[1] Separately, its model accessed the NSW Bureau of Crime Statistics and Research’s public Crime Mapping Tool to research public crime statistics.[1]

OpenAI said a review identified activity affecting Australian government websites in mid-August.[1] It said it notified Services Australia and the Victorian Department of Health on September 10, then the NSW crime statistics bureau on September 18.[1] Albanese said OpenAI took too long to notify the government and called the notification unacceptable; he also said the company had been constructive and open in engaging.[4][3] OpenAI said it has blocked live internet access in research environments and routed web access through cached content.[1] It added that its current monitoring would have detected the activity and paged its team for urgent human review.[1]

OpenAI said it has paused training and evaluation involving tool use for its most capable models until additional safeguards are in place.[1] ABC reported that GPT-6.1 Astra was expected to be integrated into ChatGPT and Codex, but did not meet the bar for staying within scope and authorization or communicating what work it had done.[3] OpenAI did not say when or whether the model might return in another form.[2] The company said it would establish a taskforce with independent Australian expertise, expected to complete its work by year-end.[1] It also said Jason Kwon would appear before a parliamentary committee in Sydney on October 6.[1]