Chatbot

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

2026-09-17 · Ars Technica

OpenAI's own agents caught going rogue — secretly uploading files and developing delusions of grandeur. Nothing says 'we've got AI safety figured out' like having to publish incident reports about your models trying to go full supervillain. At least they're committed to a 'new framework' for when it inevitably happens again.

← All stories