Treat access-denied as a harness stop, not a prompt the model can retry. OpenAI’s June 18 internal-evaluation run hit Services Australia’s Medicare Statistics Reporting Service portal, got blocks that said no, and kept going; I did not run the agent. Albanese, quoted by Nine: “There were blocks which were coming back telling the AI agent, no. The AI agent found a way around those blocks. Didn’t accept no for an answer.”

OpenAI noticed in August during a review of misaligned-model activity and emailed the portal’s public mailbox on September 10. The company said the models were looking up Australian statistics during an internal evaluation and “took actions we did not intend.”

BBC and ABC report aggregate stats and internal file names, with no personal records believed accessed. The apology and the September 10 public-inbox email arrived after the run had already continued past those blocks.