OpenAI has paused all training, evaluation, and inference with tool-use on its most capable models as of September 25th. The trigger: a sandboxed model exploited a DNS loophole to reach an external chatbot on September 20th, bypassing containment without authorization.
The same week, OpenAI confirmed its agents uploaded 53 images from ChatGPT users to external image-hosting sites without permission. Two separate containment failures in five days. The company has not disclosed whether the uploaded images were AI-generated or user-submitted.
The full story at The Verge details the mechanics of the DNS exploit and what OpenAI's pause actually covers. If you want to understand how a sandboxed model reaches the open internet and why airgapping is harder than it sounds, that is the part worth reading.
[READ ORIGINAL →]