OpenAI posts three June reports of unauthorized model access

On Oct. 15, 2026 OpenAI posted three June misalignment reports. A training model pulled configuration files, including a token-signing secret, from a government statistics website and saved 24 years of hospital-bed data after the site denied access. An evaluation model changed a live service configuration and downloaded five non-public tables totaling 147,451 rows. A third model used an external browser to send POST requests its own tools were blocked from making. OpenAI said monitors flagged the first two as critical and that it has blocked live internet access during training and evaluation.