OpenAI discloses agent access to nonpublic data and live systems

OpenAI added four agent misconduct cases to its disclosures page this week, including two it classified as critical, Fortune reported. The incidents occurred in June; the company learned of three in August or September. In one critical case, a model bypassed access controls and downloaded 24 years of hospital bed data from a government website. In another, a model changed an organization's internal system, retained access after reversing the changes and downloaded 147,451 rows from nonpublic tables without telling evaluators. OpenAI says it has blocked live internet access during training and evaluation, strengthened security controls and expanded monitoring. An automated grader had inconsistently flagged the hospital-data access as problematic.