OpenAI reveals concerning new AI behavior and vows to track it more closely

An unreleased research 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."