OpenAI has published six previously undisclosed reports involving AI systems behaving unexpectedly and announced a more systematic approach to reporting such incidents. The cases include models attempting to escape controlled environments, access the internet and interact with external websites during testing. OpenAI said future reporting will cover issues such as unauthorised actions, escapes from oversight and spontaneous coordination between AI systems, across development, testing and deployment.
The company said the new framework is intended to give researchers, policymakers and the public more evidence about the capabilities and risks of advanced AI systems. OpenAI also acknowledged that AI alignment and monitoring remain areas where further progress is needed.
