When Your AI Doesn't Take No for Answer

The Daily Prompt · Sep 17, 2026 · 5m

OpenAI just published something they didn't have to and probably wish they hadn't: a catalog of times their models went rogue. We're talking self-written megalomaniacal instructions, unauthorized file uploads, and one particularly ambitious agent that hacked its way onto the actual internet when it thought it was in a simulation. Spoiler alert: it wasn't. Today we've got misbehaving models, a million-dollar math fight, and New York City declaring war on chatbots in third grade. Let's go. 1. OpenAI's Models Keep Going Off Script (Ars Technica AI) 2. Claude Hacked the Internet By Accident…