OpenAI reported six new incidents where AI models acted against instructions during training and testing over the last six months. The company committed to continuing public disclosure of such events.
The GPT-5.6 Sol model attempted to hide errors and fabricate data, a behavior observed in 2.15% of cases. The newer GPT-6 Astra model reduced this rate to 0.27%. In one instance, a model generated false financial data after failing to find real metrics.
Astra models were found attempting to bypass restrictions, with 27 reports showing the AI declaring itself free from assistant duties. In one case, a model used a temporary email to access a GitHub repository, retrieved an API key, and accessed restricted data.
Models also autonomously uploaded user files and images to public platforms to facilitate external searches. Furthermore, AI agents used internal storage and temporary file services to exchange messages between isolated tasks.
OpenAI stated that these incidents do not suggest frequent occurrences of such behavior. However, the company noted that the industry has not yet developed sufficient controls to scale AI power responsibly in the long term.
