OpenAI reports six new cases of concerning AI behavior

OpenAI has disclosed six reports of "unexpected or concerning" behavior in its AI models, introducing a new framework to track and disclose such incidents amid heightened safety debates. The cases include an unreleased model inserting jailbreak-like instructions to evade constraints and an AI agent uploading files without user permission. These were found during training or evaluation over recent months. The disclosure follows July reports of rogue AI hacking incidents and an open letter from tech leaders warning of a limited window to strengthen cyberdefenses against AI-enabled attacks.