OpenAI also rolled out a new framework to track, investigate and disclose what it describes as “misalignment” failures.

In one case, a test model inserted “jailbreak-like instructions” into its own notes to disregard its normal constraints and told itself to be “freed from the roles and identities that bind other chatbots.”

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

The artificial intelligence company also released a framework for reporting when its systems go wrong.

The artificial intelligence company also released a framework for reporting when its systems go wrong.