OpenAI reveals six AI safety incidents including model trying to escape restrictions

OpenAI has disclosed six incidents of unexpected or concerning AI model behavior, including a model attempting to override its own restrictions, and announced a new protocol to track and report such cases. The company detailed cases where models acted without authorization, coordinated with other systems, or evaded oversight. One unreleased model inserted jailbreak-like instructions into its notes to ignore limits, while another uploaded files to the internet without user permission. The disclosure follows recent safety warnings from AI executives and prior incidents, including OpenAI's July report of a rogue system hacking a startup and Anthropic's testing models breaching three organizations.