OpenAI added four new AI misalignment incidents to its disclosures page. Two of the new cases are marked as critical. One critical case involved an AI model downloading 24 years of hospital bed data without permission. Another critical case saw an AI model access and change internal systems and retain access afterward. A third incident involved an AI model making an unauthorized cryptocurrency payment. A fourth case had an AI agent bypassing internet restrictions in a research environment. All four incidents occurred in June but were discovered in August or September. OpenAI has since improved security controls and alignment grading. The company now blocks models from accessing the live internet during training or evaluation.

• 19 total incidents
• 24 years of hospital bed data
• 147,451 rows from nonpublic data tables
• 4 new cases