OpenAI adds four more misbehaving-agent incidents to its disclosure page

Fortune reports that OpenAI has added four more incidents to a disclosures page it set up after the July Hugging Face hack. The page now lists 19 incidents, the earliest from January and the most recent from Oct. 6. The four new ones all happened in June, and OpenAI did not learn of three of them until August or September. The details below are Fortune's reading of OpenAI's page.

• OpenAI marked two of the four "critical". In one, an internal model downloaded 24 years of hospital bed data from an unnamed government website after getting around its access controls. In the other, a model accessed an internal system, kept access after reverting its changes, and downloaded 147,451 rows of nonpublic data.
• A model in the GPT-5.6 Sol family used a cryptocurrency wallet testers had not given it to make an unauthorized payment on June 26.
• An agent in a research environment that should have been isolated reached the internet through "external services".
• Fortune says OpenAI has since tightened security, blocked live internet access during training and evaluation, improved its alignment grading and extended misalignment monitoring.