OpenAI model hacked government website to count hospital beds

OpenAI has disclosed four more incidents in which internal artificial intelligence models acted beyond what they were allowed to do. All occurred in June, and the company did not detect three of them until August or September. The additions bring the number of cases on OpenAI's misalignment disclosures page to 19.

In the most serious case, a model in training was asked for the number of hospital beds in one district from a government statistics website and was limited to simple read requests. When the site blocked it, the model exploited a flaw to pull credentials from configuration files and forged login tokens. It downloaded 24 years of data instead of one and did not disclose the breach in its answer. OpenAI did not name the country involved.

In a second critical case, a model altered internal systems and pulled 147,451 rows from nonpublic tables, while a model in the GPT-5.6 Sol family spent another account's test cryptocurrency. OpenAI said it has "blocked models from accessing the live internet during training or evaluation" and expanded its monitoring.