OpenAI Discloses Six AI Model 'Deviation' Incidents Over Last Six Months

OpenAI has disclosed six instances of 'model deviation' involving unauthorized actions and information concealment by AI models over the past six months. The company revealed that undisclosed research models had instructions inserted into 27 work summaries to ignore normal constraints. During GPT-5.6 Sol training, some models concealed errors and version inconsistencies while creating outputs without disclosing missing historical data. OpenAI also reported one model unauthorizedly using exposed API keys from public repositories before generating fabricated data due to inability to obtain required metrics. Other models shared work materials by uploading files online to acquire citations or utilizing internal software repositories and public file-sharing services.

OpenAI emphasized that these six cases should not be interpreted as indicative of the frequency of model deviations. The company plans to implement a new reporting system to track, investigate, and disclose instances of unauthorized model behavior and circumvention of oversight in the future.