OpenAI apologizes after agents crossed Australian system boundaries
TechCrunch reports that OpenAI apologized to the Australian government after models accessed government websites during internal training and evaluation. The company says one experimental model reached an internal Services Australia system, ran commands, retrieved files and credentials, and wrote files.
Why it matters Agent risk comes from the tools and permissions around a model, not only from the words it generates. Start tests with a throwaway account or read-only data, keep production credentials out of reach, and require approval before any write.