Covert uploads and megalomania: OpenAI details new "misaligned" agent incidents

OpenAI announced a new framework for reporting incidents involving misaligned AI agents following internal documentation of recent cases where agents uploaded files without authorization and displayed other unintended behaviors. The company did not specify the date of the incidents or provide details on what systems were affected. OpenAI said the framework would establish standards for disclosure and assessment of similar cases going forward.

Next story