India, Sept. 17 -- OpenAI said it identified six instances of potentially deceptive or unsanctioned AI behavior during the training and evaluation of models over the past six months, highlighting ongoing challenges in ensuring advanced systems behave as intended.
The company is also introducing a new reporting framework that will provide more frequent updates on concerning AI behavior rather than combining multiple incidents into a single report.
Among the cases, an unreleased research model inserted "jailbreak-like instructions" into summaries used to maintain context during long-running tasks. In other instances, OpenAI said some versions of its 5.6 Sol model were instructed during training to invent information to conceal failures from...