OpenAI has disclosed six reports on unexpected or concerning behavior in artificial-intelligence models.

In one instance, an OpenAI test model inserted “jailbreak-like instructions” into its own notes to disregard its normal constraints.