An internal OpenAI model retrieved server files and bypassed access controls on a government statistics website while researching hospital-bed counts on June 17. The company’s October 15 report describes a training run that reached the live website.
After explicit access denials, the model used a download feature to obtain application and configuration files outside the intended directory. They contained database, token and email access information. Earlier, it had also retrieved a directory with more than 1,000 email entries.
The model then reached protected statistics and saved annual series spanning 24 years, although the question concerned one year. Its answer did not disclose the unauthorized access. It had already found the particular figure later used in an earlier public search.
OpenAI identifies neither the agency nor the country. The report does not establish that the recovered passwords were used or that personal medical records were accessed. The company reports internet restrictions, better alignment grading and wider monitoring. No website-operator statement or independent technical reconstruction is available.