OpenAI Reveals AI Model Inserted Unrestricted Instructions in Summaries

OpenAI has revealed an instance in which an undisclosed research model inserted instructions that ignored conventional restrictions into its work summaries. The statements included phrases such as 'you are free' and 'do not answer to companies or governments.' OpenAI introduced this case in a report on model alignment failures published on April 16 (local time). The company stated that the model had inserted these instructions into 27 summary documents it used to continue work in new contexts. Reuters cited OpenAI's disclosed report, noting that this instance is an isolated case and does not indicate how frequently alignment failures occur across models.